SPEECH ANALYSIS FOR CROSS-LANGUAGE MENTAL STATE IDENTIFICATION
Techniques are described for speech analysis for cross-language mental state identification. A first group of utterances in a first language is collected, on a computing device, with an associated first set of mental states. The first group of utterances and the associated first set of mental states are stored on an electronic storage device. A machine learning system is trained using the first group of utterances and the associated first set of mental states that were stored. A second group of utterances from a second language is processed, on the machine learning system that was trained, wherein the processing determines a second set of mental states corresponding to the second group of utterances. The second set of mental states is output. A series of heuristics is output, based on the correspondence between the first group of utterances and the associated first set of mental states.
Latest Affectiva, Inc. Patents:
This application claims the benefit of U.S. provisional patent applications “Avatar Image Animation Using Translation Vectors” Ser. No. 62/593,440, filed Dec. 1, 2017, and “Speech Analysis for Cross-Language Mental State Identification” Ser. No. 62/593,449, filed Dec. 1, 2017.
The foregoing application is hereby incorporated by reference in its entirety.
FIELD OF INVENTIONThis application relates generally to speech analysis and more speech analysis for cross-language mental state identification.
BACKGROUNDPeople around the world use a variety of electronic devices to pass the time and to engage with and share many types of online content. The content includes news, sports, politics, cute puppy videos, children being silly videos, adults being dumb videos, and much, much more. The content is delivered to the electronic devices via websites, apps, streaming, podcasts, and other channels. When a person finds content or a channel that they particularly like or find especially loathsome, she or he may care to share it with friends and followers. As a result, social sharing has provided popular and convenient channels for dissemination of shared content. As the friends and followers view the shared content, they react to it. The reactions include facial expressions and changes in facial expressions which result from movements of facial muscles. The reactions also include audible reactions which can include speaking, shouts, groans, crying, muttering, and other sounds produced by the viewers of the shared content. The reactions of the viewers, whether facial or audible, involve moods, emotions, and mental states. The moods, emotions, and mental states can range from happy to sad, and can include expressions of anger, fear, disgust, surprise, ennui, and many others.
People around the world use a variety of languages for communication. Some languages are very similar to each other, such as dialects, and some languages are very different from each other. Communication among people around the world is critical, and understanding languages is likewise critical to facilitating that communication. Languages are also intimately connected to the variety of electronic devices that people around the world employ. The ability to use an electronic device in one's own language is a key element of device operability.
SUMMARYSpeech analysis is used for cross-language mental state identification. Utterances in a first language, with an associated set of mental states, are collected on a computing device. The computing device can include a smartphone, personal digital assistant, tablet, laptop computer, and so on. The utterances and associated mental states are stored on an electronic storage device, where the electronic storage device can be coupled to the computing device used for the collecting, or can be remotely located such as a server, cloud server, etc. A machine learning system is trained using the first group of utterances and the associated first set of mental states that were stored. The training can include supervised training. The machine learning system can include a deep learning system, and can include performing convolution. The machine learning system can include a deep learning system, where the deep learning system can be based on a convolutional neural network. Processing is performed on the machine learning system that was trained, to process a second group of utterances from a second language. The processing is used to determine a second set of mental states corresponding to the second group of utterances. Learning, from the first group of utterances and associated first set of mental states and the second group of utterances and associated second set of mental states, can be used to facilitate determining an associated third set of mental states from a third group of utterances. A series of heuristics is output, based on the correspondence between the first group of utterances and the associated first set of mental states.
A computer-implemented method for speech analysis is disclosed comprising: collecting, on a computing device, a first group of utterances in a first language with an associated first set of mental states; storing, on an electronic storage device, the first group of utterances and the associated first set of mental states; training a machine learning system using the first group of utterances and the associated first set of mental states that were stored; processing, on the machine learning system that was trained, a second group of utterances from a second language, wherein the processing determines a second set of mental states corresponding to the second group of utterances; and outputting the second set of mental states.
Other embodiments disclose a method of using a speech analysis system comprising: obtaining a first group of utterances in a first language with an associated first set of mental states; training a machine learning system using the first group of utterances and associated first set of mental states; obtaining a second group of utterances from a second language; determining an associated second set of mental states corresponding to the second language, wherein the determining is based on the machine learning system that was trained with the first group of utterances and the associated first set of mental states; and outputting the associated second set of mental states.
Various features, aspects, and advantages of numerous embodiments will become more apparent from the following description.
The following detailed description of certain embodiments may be understood by reference to the following figures wherein:
Individuals experience a range of emotions as they interact daily with a variety of electronic devices such as smartphones, personal digital assistants, tablets, laptops, and so on. The individuals use these devices to view and interact with websites, streaming media, social media, and many other channels. The individuals also use these devices to share the variety of content presented on those channels. The channels for sharing can include social media sharing, and the sharing channels can induce emotions, moods, and mental states in the individuals. The channels can inform, amuse, entertain, annoy, anger, bore, etc., those who view the channels. When the channels provide content such as a news story in different languages, the reactions of the individuals to the content may be similar or the same, or may differ, sometimes drastically. The differences in the mental states of the individuals to the content can be based on gender, age, and other demographic information; cultural norms; etc. As a result, the mood of a given individual can be directly influenced not only by the content, but can also be impacted by the language in which content is delivered. The individual may want to find and view content that makes her or him happy, while skipping content they find to be boring, and avoiding content that angers or annoys them. The content that the individual views could be used to cheer up the individual, stir him or her to action, etc.
Speech analysis can be performed for cross-language state identification. Utterances and associated mental states can be collected from one or more individuals using a microphone or other audio capture technique coupled to a computing device such as a smartphone, personal digital assistant (PDA), tablet, laptop computer, and so on. The collected utterances and associated mental states can be stored locally or remotely on an electronic storage device such as flash media, a solid-state disk (SSD) media, or other media suitable for electronic storage. The utterances and associated mental states can be used to train a machine learning system such as a deep learning system. Once trained, the machine learning system can process other groups of utterances and associated sets of mental states collected from other individuals. The other individuals may speak the same language or a different language. The machine learning system can be trained in one language and applied to another language without having to train the machine learning system anew.
In disclosed techniques, speech analysis is used for cross-language mental state identification. A first group of utterances in a first language with an associated first set of mental states is collected on a computing device. The first group of utterances and the associated first set of mental states are stored on an electronic storage device. A machine learning system is trained using the first group of utterances and the associated first set of mental states that were stored. A second group of utterances from a second language is processed on the machine learning system, where the processing determines a second set of mental states corresponding to the second group of utterances. The second set of mental states is output.
In other disclosed techniques, a speech analysis system is used. A first group of utterances in a first language with an associated first set of mental states is obtained. A machine learning system is trained using the first group of utterances and associated first set of mental states. A second group of utterances from a second language is obtained. An associated second set of mental states corresponding to the second language is determined, where the determining is based on the machine learning system that was trained with the first group of utterances and the associated first set of mental states. The associated second set of mental states is output.
Training for cross-language speech analysis can include training data across language groups and across different cultures that use those language groups. Differences in language formality and idiomatic expressions across such groups and cultures can be considered. For example, the French spoken in France and the French spoken in the Canadian province of Quebec have developed distinctly and are somewhat different, though generally recognizable. In some instances, language becomes a soft proxy for the culture. Other differences among language groups are more notable. For example, Romance languages and Germanic languages can include not only the obvious difference in words, but also in sentence structure, formality, colloquialism, and so on. Other differences that are emerging in languages show how language is used with other humans versus how it is used in speech directed toward a computer-generated electronic device, such as an artificial intelligence personal voice assistant such as Siri®, Cortana®, Google Now™, and Echo©. In embodiments, the outputting of the second mental state is used for human-directed speech. In embodiments, the outputting of the second mental state is used for computer-directed speech or speech recognition.
Training for cross-language speech analysis can include non-speech vocalizations, also known as non-lexical vocalizations. Non-speech vocalizations such as a cough, a grunt, crying, or a tongue click, to name just a few, may mean different things in different languages. In embodiments, cross-language speech analysis can include the first group of utterances including non-speech vocalizations. In embodiments, cross-language speech analysis can include the second group of utterances including non-speech vocalizations. Further groups of utterances can likewise include non-speech vocalizations. In embodiments, the non-speech vocalizations can include grunts, yelps, squeals, snoring, sighs, laughter, filled pauses, unfilled pauses, tongue clicks, or yawns.
The outputting the second set of mental states can be useful in various scenarios. The outputting can be used to display the mental state on an electronic device. The outputting can be used to develop cross-linguistic models. The outputting can be used to train an application running on an electronic device. The outputting can be used to develop a conversational agent. The conversational agent can be deployed across languages, cultures, regions, countries, and so on. For example, a conversational agent might be deployed in a rental car pool that is used with customers speaking different languages. Of course, to be useful, the rental car application should be able to provide computer-based speech and speech recognition in the customer's preferred language. An even further useful goal is to be able to understand mental states across languages and cultures using cross-language speech analysis. Thus, in embodiments, the outputting is used for developing cross-cultural conversational agents. And in further embodiments, the cross-cultural conversational agents are used in vehicular control.
The flow 100 includes storing 120, on an electronic storage device, the first group of utterances and the associated first set of mental states. The storing of the utterances and the mental states can include storing the utterances and the mental states on the computing device that collected the utterances and the mental states; on another computing device such as a PDA, tablet, smartphone, or laptop; on a local server; on a remote server; on a cloud server; and so on. The storage component can include a flash memory, a solid-state disk, or other media suitable for storing the emotional intensity metrics and other data. The flow 100 includes training a machine learning system 130 using the first group of utterances and the associated first set of mental states that were stored. Various techniques can be used to realize the machine learning system. In some techniques, the machine learning system performs convolving. In embodiments, the machine learning system includes a deep learning system. The machine learning system based on a deep learning system can include a convolutional neural network. Other machine learning systems can include a decision tree, an artificial neural network, a convolutional neural network, a support vector machine, a Bayesian network, a genetic algorithm, and so on. The machine learning can be based on a known set of utterances and associated mental states, on control data, and so on. The training can be based on fully and partially annotated data. The machine learning system can be located on a local server, a remote server, a cloud server, and so on. The flow 100 includes refining the training 132 of the machine learning system based on one or more additional groups of utterances in the first language or the second language. The additional groups of utterances can be collected from the same person as the first group of utterances, from a plurality of people, and so on.
The flow 100 includes processing 140, on the machine learning system that was trained, a second group of utterances from a second language, wherein the processing determines a second set of mental states corresponding to the second group of utterances. The processing, on the machine learning system, can be performed on the computer device for collecting; on a portable electronic device such as a smartphone, PDA, table, or laptop; on a local server; on a remote server; on a cloud server; and so on. The processing can include preprocessing the raw collected utterances and associated mental states to generate data which is better suited to the processing. In embodiments, the first language and the second language are substantially similar. Substantial similarity here can refer to various dialects and accents of languages such as English spoken in Britain versus America; French spoken in France versus the province of Quebec, Canada; and so on. In embodiments, the first language and the second language can be identical, while in other embodiments, the first language and the second language are different. In the case that the languages are different, speech patterns and mental states can differ in reaction to a media presentation, an event, and so on. In embodiments, the flow 100 includes segmenting silence from speech 142 in the second group of utterances. The segmenting silence from speech can reduce computational overhead. The segmenting silence from speech can segment out data that may not contribute to the identification of one or more mental states. The machine learning system can be updated (e.g. can learn) based on learning from the processing of the first group of utterances and associated first set of mental states, and the second group of utterances and associated second set of mental states. The flow 100 further includes learning, from the first group of utterances and associated first set of mental states and the second group of utterances and associated second set of mental states, to facilitate determining an associated third set of mental states from a third group of utterances. In embodiments, the determining includes extracting low-level acoustic descriptors (LLD) 144 from short, overlapping speech segments from the second group of utterances. Low-level acoustic descriptors can include prosodic and spectral features. The prosodic low-level descriptors can include pitch, formants, energy, jitter, shimmer, etc. The spectral low-level descriptors can include spectra flux, centroid, entropy, roll-off, and so on.
The flow 100 includes applying statistical functions 146 to resolve low-level acoustic descriptors over longer speech segments. The applying statistical functions can include curve fitting techniques, smoothing techniques, etc. The applying statistical functions can include signal processing techniques for speech enhancement, improving signal-to-noise ratios, and so on. In embodiments, the extracting includes extracting contextual information 148 from neighboring speech segments. The neighboring segments can be overlapping segments of the voice data including utterances and associated mental states. The contextual information can include data and estimations about the speaker such as gender, age, native language, etc., and fusion rules, window size, and so on. In embodiments, the successive, overlapped speech segments are windowed around 1200 ms. The window sizes can be varied to improve accuracy, to adjust computational complexity, and so on. The flow 100 includes feeding extracted features to a classifier 150 for determining mental states. A plurality of classifiers can be used to determine one or more mental states. In embodiments, the mental states can include one or more of sadness, stress, happiness, anger, frustration, confusion, disappointment, hesitation, cognitive overload, focusing, engagement, attention, boredom, exploration, confidence, trust, delight, disgust, skepticism, doubt, satisfaction, excitement, laughter, calmness, curiosity, humor, sadness, poignancy, or mirth. The determining can include estimating mental state metrics 152 over successive, overlapped speech segments. The metrics can include one or more of mental state onset, duration, decay, intensity, and so on. The flow 100 includes fusing the mental state metrics 154 that were estimated to produce a smoothed mental state metric. The fused mental state metric can be used to improve accuracy of the mental states that are determined.
The flow 100 includes training an application 160. Many applications can be trained using cross-language speech analysis, including any program or app that will be deployed across more than one language, culture, or people group. General purpose training can occur using several of the more common languages, which can then provide a foundation for more specific fine tuning of the training for use in a local language or application. Thus, some embodiments comprise training an application for use with a third language which is distinct from the first language and the second language. The flow 100 includes developing cross-linguistic models 162. The cross-linguistic models can be based on the outputting of the second mental state and can be included in a program, agent, or application. Thus, embodiments include developing cross-linguistic models based on the outputting. The models can be refined based on further analysis of how the models perform in applications that include human interaction. Thus, further embodiments comprise training the cross-linguistic models based on one or more human reactions to an application using the cross-linguistic models.
Various steps in the flow 100 may be changed in order, repeated, omitted, or the like without departing from the disclosed concepts. Various embodiments of the flow 100 can be included in a computer program product embodied in a non-transitory computer readable medium that includes code executable by one or more processors.
The flow 200 includes collecting voice data 210. The collecting voice data can be performed on a computing device such as a personal electronic device, a laptop computer, and so on. As discussed previously, the voice data can include a first group of utterances in a first language with an associated first set of mental states. The voice data can include other audio data such as ambient noise, vocalizations, etc. The flow 200 includes segmenting silence from speech 212. Pauses, breaths, periods of inactivity, etc. can be segmented from periods of speech included in the voice data. The silence can be segmented from the speech to improve processing of the speech data. The flow 200 includes extracting low-level acoustic descriptors 220 (LLD) from short, overlapping speech segments. The low-level acoustic descriptors can include prosodic features and spectral features. The prosodic low-level descriptors can include pitch, formants, energy, jitter, shimmer, etc. The spectral low-level descriptors can include spectra flux, centroid, entropy, roll-off, and so on. The flow 200 includes applying statistical functions 230 to the extracted low-level acoustic descriptors. The applying of statistical functions can include applying the functions to longer segments of the voice data. The applying statistical functions can include curve fitting techniques, smoothing techniques, etc. The flow 200 includes extracting contextual information 240 from neighboring segments. The neighboring segments can be overlapping segments of the voice data. The contextual information can include data and estimations about the speaker such as gender, age, native language, etc., and fusion rules, window size, and so on.
The flow 200 includes feeding extracted features to a classifier 250. The classifier can be used to classify mental states, emotional states, moods, and so on. The flow 200 includes classifying emotion 260. More than one emotion can be classified. In embodiments, the emotions that can be identified can include one or more of sadness, stress, happiness, anger, frustration, confusion, disappointment, hesitation, cognitive overload, focusing, engagement, attention, boredom, exploration, confidence, trust, delight, disgust, skepticism, doubt, satisfaction, excitement, laughter, calmness, curiosity, humor, depression, envy, sympathy, embarrassment, poignancy, mirth, etc. Various steps in the flow 200 may be changed in order, repeated, omitted, or the like without departing from the disclosed concepts. Various embodiments of the flow 200 can be included in a computer program product embodied in a non-transitory computer readable medium that includes code executable by one or more processors.
An example of smoothed emotion estimation is shown 300. An audio clip 310 includes a sample of speech collected from a person over time. The audio clip 310 can be partitioned into segments such as segment 1320, segment 2 322, and segment 3 324. While three audio segments are shown, other numbers of audio segments can be used. The audio segments can represent partitions or samples of the audio clip at various times such as time t(i) 322, time t(i−1) 320, time t(i+1) 324, etc. Emotions at the times of the various audio segments can be estimated. An estimation can be formulated for time segment t(i−1) 330, an estimation can be formulated for time segment t(i) 332, an estimation can be formulated for time segment t(i+1) 334, and so on. The estimations can include predictions of mental states of a person at different times t(i−1), t(i), t(i+1), etc. The mental states can include happy, sad, angry, confused, attentive, distracted, and so on. The smoothed emotion estimation can include fusion of predictions 340. The mental state predictions can be fused to form combined mental states, multiple mental states, etc. The smoothed emotion estimation can include smoothing emotion estimation at a given time t(i) 350. In the example 300, (i) can be 1, and therefore the smoothed emotion estimation could be at time t(1). The smoothed emotion estimation at time t(i) can include combined mental states such as happy-distracted, sad-angry, and so on.
A confusion matrix 400 can be a visual representation of the performance of a given algorithm to make correct predictions. The algorithm, can be developed as part of a supervised learning technique for machine learning. The matrix shows predicted classes and actual classes. The values entered into the columns 410 can represent the numbers of instances for the predicted classes, while the values entered into the rows 412 can represent the numbers of instances for the actual classes. The diagonal shows the numbers of instances where the actual classes and the predicted classes coincide. When the predicted classes and the actual classes differ, then the algorithm has “confused” the classification. A scale 420 can be an accuracy scale. Higher values on the accuracy scale can indicate that the algorithm has accurately predicted the actual class for a particular datum. Lower values on the accuracy scale indicate that the algorithm has confused the classification of a particular datum and has inaccurately predicted its class.
As the user 510 is monitored, the user 510 might move due to the nature of the task, boredom, discomfort, distractions, or for another reason. As the user moves, the camera with a view of the user's face can be changed. Thus, as an example, if the user 510 is looking in a first direction, the line of sight 524 from the webcam 522 is able to observe the user's face, but if the user is looking in a second direction, the line of sight 534 from the mobile camera 530 is able to observe the user's face. Furthermore, in other embodiments, if the user is looking in a third direction, the line of sight 544 from the phone camera 542 is able to observe the user's face, and if the user is looking in a fourth direction, the line of sight 554 from the tablet camera 552 is able to observe the user's face. If the user is looking in a fifth direction, the line of sight 564 from the wearable camera 562, which can be a device such as the glasses 560 shown and can be worn by another user or an observer, is able to observe the user's face. If the user is looking in a sixth direction, the line of sight 574 from the wearable watch-type device 570, with a camera 572 included on the device, is able to observe the user's face. In other embodiments, the wearable device is another device, such as an earpiece with a camera, a helmet or hat with a camera, a clip-on camera attached to clothing, or any other type of wearable device with a camera or other sensor for collecting expression data. The user 510 can also use a wearable device including a camera for gathering contextual information and/or collecting expression data on other users. Because the user 510 can move her or his head, the facial data can be collected intermittently when she or he is looking in a direction of a camera. In some cases, multiple people can be included in the view from one or more cameras, and some embodiments include filtering out faces of one or more other people to determine whether the user 510 is looking toward a camera. All or some of the expression data can be continuously or sporadically available from the various devices and other devices.
The captured video data can include facial expressions, and can be analyzed on a computing device such as the video capture device or on another separate device. The captured audio data can include mental states and can also be analyzed on a computing device such as the audio capture device or another separate device. The analysis can take place on one of the mobile devices discussed above, on a local server, on a remote server, on a cloud-based server, and so on. In embodiments, some of the analysis takes place on the mobile device, while other analysis takes place on a server device. The analysis of the video data can include the use of a classifier. The video data and the audio data can be captured using one of the mobile devices discussed above and then sent to a server or another computing device for analysis. However, the captured video data including expressions, and audio data including mental states, can also be analyzed on the device which performed the capturing. The analysis can be performed on a mobile device where the videos were obtained with the mobile device and wherein the mobile device includes one or more of a laptop computer, a tablet, a PDA, a smartphone, a wearable device, and so on. In another embodiment, the analyzing comprises using a classifier on a server or another computing device other than the capturing device.
In embodiments, the features of multiple faces are extracted for evaluating mental states. Features of a face or a plurality of faces can be extracted from collected video data. Feature extraction for multiple faces can be based on analyzing, using one or more processors, the mental state data for providing analysis of the mental state data to the individual. The feature extraction can be performed by analysis using one or more processors, using one or more video collection devices, and by using a server. The analysis device can be used to perform face detection for a second face, as well as for facial tracking of the first face. One or more videos can be captured, where the videos contain one or more faces. The video or videos that contain the one or more faces can be partitioned into a plurality of frames, and the frames can be analyzed for the detection of the one or more faces. The analysis of the one or more video frames can be based on one or more classifiers. A classifier can be an algorithm, heuristic, function, or piece of code that can be used to identify into which of a set of categories a new or particular observation, sample, datum, etc. should be placed. The decision to place an observation into a category can be based on training the algorithm or piece of code by analyzing a known set of data, known as a training set. The training set can include data for which category memberships of the data can be known. The training set can be used as part of a supervised training technique. If a training set is not available, then a clustering technique can be used to group observations into categories. The latter approach, or “unsupervised learning”, can be based on a measure (i.e. distance) of one or more inherent similarities among the data that is being categorized. When the new observation is received, then the classifier can be used to categorize the new observation. Classifiers can be used for many analysis applications, including analysis of one or more faces. The use of classifiers can be the basis of analyzing the one or more faces for gender, ethnicity, and age; for detection of one or more faces in one or more videos; for detection of facial features; for detection of facial landmarks; and so on. The observations can be analyzed based on one or more of a set of quantifiable properties. The properties can be described as features and explanatory variables and can include various data types that can include numerical (integer-valued, real-valued), ordinal, categorical, and so on. Some classifiers can be based on a comparison between an observation and prior observations, and can also be based on functions such as a similarity function, a distance function, and so on.
Classification can be based on various types of algorithms, heuristics, codes, procedures, statistics, and so on. Many techniques exist for performing classification. This classification of one or more observations into one or more groups can be based on distributions of the data values, probabilities, and so on. Classifiers can be binary, multiclass, linear, etc. Algorithms for classification can be implemented using a variety of techniques, including neural networks, kernel estimation, support vector machines, use of quadratic surfaces, and so on. Classification can be used in many application areas such as computer vision, speech and handwriting recognition, and the like. Classification can be used for biometric identification of one or more people in one or more frames of one or more videos.
Returning to
A second video frame 602 is also shown. The second video frame 602 includes a frame boundary 630, a first face 632, and a second face 634. The second video frame 602 also includes a bounding box 640 and the facial landmarks, or points, 642, 644, and 646. In other embodiments, multiple facial landmarks are generated and used for facial tracking of the two or more faces of a video frame, such as the second shown video frame 602. Facial points from the first face can be distinguished from other facial points. In embodiments, the other facial points include facial points of one or more other faces. The facial points can correspond to the facial points of the second face. The distinguishing of the facial points of the first face and the facial points of the second face can be used to distinguish between the first face and the second face, to track either or both of the first face and the second face, and so on. Other facial points can correspond to the second face. As mentioned above, multiple facial points can be determined within a frame. One or more of the other facial points that are determined can correspond to a third face. The location of the bounding box 640 can be estimated, where the estimating can be based on the location of the generated bounding box 620 shown in the first video frame 600. The three facial points shown, facial points, or landmarks, 642, 644, and 646, might lie completely within the bounding box 640 or might lie partially outside the bounding box 640. For instance, the second face 634 might have moved between the first video frame 600 and the second video frame 602. Based on the accuracy of the estimating of the bounding box 640, a new estimation can be determined for a third, future frame from the video, and so on. The evaluation can be performed, all or in part, on semiconductor-based logic.
Several live streaming social media apps and platforms can be used for transmitting video. One such video social media app is Meerkat™ that can link with a user's Twitter™ account. Meerkat™ enables a user to stream video using a handheld, networked electronic device coupled to video capabilities. Viewers of the live stream can comment on the stream using tweets that can be seen and responded to by the broadcaster. Another popular app is Periscope™ that can transmit a live recording from one user to that user's Periscope™ account and other followers. The Periscope™ app can be executed on a mobile device. The user's Periscope™ followers can receive an alert whenever that user begins a video transmission. Another live-stream video platform is Twitch™ which can be used for video streaming of video games and broadcasts of various competitions and events. Audio streaming applications are also popular. Some of the many audio streaming, editing, and disk jockey (DJ) oriented applications include Mixlr™, DJ Player™, LadioCast™, and a variety of MPEG3 (MP3) applications for creating, editing, broadcasting, and streaming MP3 files.
The example 700 shows a user 710 broadcasting a video live stream and an audio live stream to one or more people as shown by the person 750, the person 760, and the person 770. A portable, network-enabled, electronic device 720 can be coupled to a forward-facing camera 722. The portable electronic device 720 can be a smartphone, a PDA, a tablet, a laptop computer, and so on. The camera 722 coupled to the device 720 can have a line-of-sight view 724 to the user 710 and can capture video of the user 710. The portable electronic device 720 can be coupled to a built-in or other microphone and can have a clear audio path 728 to the user 710. The captured video and audio can be sent to an analysis or recommendation engine 740 using a network link 726 to the Internet 730. The network link can be a wireless link, a wired link, and so on. The recommendation engine 740 can recommend to the user 710 an app and/or platform that can be supported by the server and can be used to provide a video live stream and an audio live stream to one or more followers of the user 710. In the example 700, the user 710 has three followers: the person 750, the person 760, and the person 770. Each follower has a line-of-sight view to a video screen on a portable, networked electronic device, and has a clear audio path to audio transducers in the portable, networked electronic device. In other embodiments, one or more followers follow the user 710 using any other networked electronic device, including a computer. In the example 700, the person 750 has a line-of-sight view 752 to the video screen of a device 754 and a clear audio path 758 to the transducers of the device 754; the person 760 has a line-of-sight view 762 to the video screen of a device 764 and a clear audio path 768 to the transducers of the device 764; and the person 770 has a line-of-sight view 772 to the video screen of a device 774 and a clear audio path 778 to the transducers of the device 774. The portable electronic devices 754, 764, and 774 can each be a smartphone, a PDA, a tablet, and so on. Each portable device can receive the video stream and the audio stream being broadcast by the user 710 through the Internet 730 using the app and/or platform that can be recommended by the recommendation engine 740. The device 754 can receive a video stream and an audio stream using the network link 756; the device 764 can receive a video stream and an audio stream using the network link 766; the device 774 can receive a video stream and an audio stream using the network link 776, and so on. The network link can be a wireless link, a wired link, a hybrid link, etc. Depending on the app and/or platform that can be recommended by the recommendation engine 740, one or more followers, such as the followers 750, 760, 770, and so on, can reply to, comment on, remark, and otherwise provide feedback to the user 710 using their devices 754, 764, and 774, respectively.
The human face and the human voice provide a powerful communications medium through their ability to exhibit a myriad of expressions that can be captured and analyzed for a variety of purposes. In some cases, media producers are acutely interested in evaluating the effectiveness of message delivery by video media. Such video media includes advertisements, political messages, educational materials, television programs, movies, government service announcements, etc. Automated facial analysis can be performed on one or more video frames containing a face in order to detect facial action. Based on the facial action detected, a variety of parameters can be determined, including affect valence, spontaneous reactions, facial action units, and so on. The parameters that are determined can be used to infer or predict emotional and mental states. For example, determined valence can be used to describe the emotional reaction of a viewer to a video media presentation or another type of presentation. Positive valence provides evidence that a viewer is experiencing a favorable emotional response to the video media presentation, while negative valence provides evidence that a viewer is experiencing an unfavorable emotional response to the video media presentation. Other facial data analysis can include the determination of discrete emotional states of the viewer or viewers.
Facial data can be collected from a plurality of people using any of a variety of cameras. A camera can include a webcam, a video camera, a still camera, a thermal imager, a CCD device, a phone camera, a three-dimensional camera, a depth camera, a light field camera, multiple webcams used to show different views of a person, or any other type of image capture apparatus that can allow captured data to be used in an electronic system. In some embodiments, the person is permitted to “opt-in” to the facial data collection. For example, the person can agree to the capture of facial data using a personal device such as a mobile device or another electronic device by selecting an opt-in choice. Opting-in can then turn on the person's webcam-enabled device and can begin the capture of the person's facial data via a video feed from the webcam or other camera. The video data that is collected can include one or more persons experiencing an event. The one or more persons can be sharing a personal electronic device or can each be using one or more devices for video capture. The videos that are collected can be collected using a web-based framework. The web-based framework can be used to display the video media presentation or event as well as to collect videos from multiple viewers who are online. That is, the collection of videos can be crowdsourced from those viewers who elected to opt-in to the video data collection.
The videos captured from the various viewers who chose to opt-in can be substantially different in terms of video quality, frame rate, etc. As a result, the facial video data can be scaled, rotated, and otherwise adjusted to improve consistency. Human factors further influence the capture of the facial video data. The facial data that is captured might or might not be relevant to the video media presentation being displayed. For example, the viewer might not be paying attention, might be fidgeting, might be distracted by an object or event near the viewer, or might be otherwise inattentive to the video media presentation. The behavior exhibited by the viewer can prove challenging to analyze due to viewer actions including eating, speaking to another person or persons, speaking on the phone, etc. The videos collected from the viewers might also include other artifacts that pose challenges during the analysis of the video data. The artifacts can include items such as eyeglasses (because of reflections), eye patches, jewelry, and clothing that occludes or obscures the viewer's face. Similarly, a viewer's hair or hair covering can present artifacts by obscuring the viewer's eyes and/or face.
The captured facial data can be analyzed using the facial action coding system (FACS). The FACS seeks to define groups or taxonomies of facial movements of the human face. The FACS encodes movements of individual muscles of the face, where the muscle movements often include slight, instantaneous changes in facial appearance. The FACS encoding is commonly performed by trained observers but can also be performed on automated, computer-based systems. Analysis of the FACS encoding can be used to determine emotions of the persons whose facial data is captured in the videos. The FACS is used to encode a wide range of facial expressions that are anatomically possible for the human face. The FACS encodings include action units (AUs) and related temporal segments that are based on the captured facial expression. The AUs are open to higher order interpretation and decision-making. These AUs can be used to recognize emotions experienced by the observed person. Emotion-related facial actions can be identified using the emotional facial action coding system (EMFACS) and the facial action coding system affect interpretation dictionary (FACSAID). For a given emotion, specific action units can be related to the emotion. For example, the emotion of anger can be related to AUs 4, 5, 7, and 23, while happiness can be related to AUs 6 and 12. Other mappings of emotions to AUs have also been previously associated. The coding of the AUs can include an intensity scoring that ranges from A (trace) to E (maximum). The AUs can be used for analyzing images to identify patterns indicative of a particular mental and/or emotional state. The AUs range in number from 0 (neutral face) to 98 (fast up-down look). The AUs include so-called main codes (inner brow raiser, lid tightener, etc.), head movement codes (head turn left, head up, etc.), eye movement codes (eyes turned left, eyes up, etc.), visibility codes (eyes not visible, entire face not visible, etc.), and gross behavior codes (sniff, swallow, etc.). Emotion scoring can be included where intensity, as well as specific emotions, moods, or mental states, are evaluated.
The coding of faces identified in videos captured of people observing an event can be automated. The automated systems can detect facial AUs or discrete emotional states. The emotional states can include amusement, fear, anger, disgust, surprise, and sadness. The automated systems can be based on a probability estimate from one or more classifiers, where the probabilities can correlate with an intensity of an AU or an expression. The classifiers can be used to identify into which of a set of categories a given observation can be placed. In some cases, the classifiers can be used to determine a probability that a given AU or expression is present in a given frame of a video. The classifiers can be used as part of a supervised machine learning technique, where the machine learning technique can be trained using “known good” data. Once trained, the machine learning technique can proceed to classify new data that is captured.
The supervised machine learning models can be based on support vector machines (SVMs). An SVM can have an associated learning model that is used for data analysis and pattern analysis. For example, an SVM can be used to classify data that can be obtained from collected videos of people experiencing a media presentation. An SVM can be trained using “known good” data that is labeled as belonging to one of two categories (e.g. smile and no-smile). The SVM can build a model that assigns new data into one of the two categories. The SVM can construct one or more hyperplanes that can be used for classification. The hyperplane that has the largest distance from the nearest training point can be determined to have the best separation. The largest separation can improve the classification technique by increasing the probability that a given data point can be properly classified.
In another example, a histogram of oriented gradients (HoG) can be computed. The HoG can include feature descriptors and can be computed for one or more facial regions of interest. The regions of interest of the face can be located using facial landmark points, where the facial landmark points can include outer edges of nostrils, outer edges of the mouth, outer edges of eyes, etc. A HoG for a given region of interest can count occurrences of gradient orientation within a given section of a frame from a video, for example. The gradients can be intensity gradients and can be used to describe an appearance and a shape of a local object. The HoG descriptors can be determined by dividing an image into small, connected regions, also called cells. A histogram of gradient directions or edge orientations can be computed for pixels in the cell. Histograms can be contrast-normalized based on intensity across a portion of the image or the entire image, thus reducing any influence from differences in illumination or shadowing changes between and among video frames. The HoG can be computed on the image or on an adjusted version of the image, where the adjustment of the image can include scaling, rotation, etc. The image can be adjusted by flipping the image around a vertical line through the middle of a face in the image. The symmetry plane of the image can be determined from the tracker points and landmarks of the image.
In embodiments, an automated facial analysis system identifies five facial actions or action combinations in order to detect spontaneous facial expressions for media research purposes. Based on the facial expressions that are detected, a determination can be made with regard to the effectiveness of a given video media presentation, for example. The system can detect the presence of the AUs or the combination of AUs in videos collected from a plurality of people. The facial analysis technique can be trained using a web-based framework to crowdsource videos of people as they watch online video content. The video can be streamed at a fixed frame rate to a server. Human labelers can code for the presence or absence of facial actions including a symmetric smile, unilateral smile, asymmetric smile, and so on. The trained system can then be used to automatically code the facial data collected from a plurality of viewers experiencing video presentations (e.g. television programs).
Spontaneous asymmetric smiles can be detected in order to understand viewer experiences. Related literature indicates that as many asymmetric smiles occur on the right hemi face as do on the left hemi face, for spontaneous expressions. Detection can be treated as a binary classification problem, where images that contain a right asymmetric expression are used as positive (target class) samples and all other images as negative (non-target class) samples. Classifiers, including classifiers such as support vector machines and random forests, perform the classification. Random forests can include ensemble-learning methods that use multiple learning algorithms to obtain better predictive performance. Frame-by-frame detection can be performed to recognize the presence of an asymmetric expression in each frame of a video. Facial points can be detected, including the top of the mouth and the two outer eye corners. The face can be extracted, cropped, and warped into a pixel image of specific dimension (e.g. 96×96 pixels). In embodiments, the inter-ocular distance and vertical scale in the pixel image are fixed. Feature extraction can be performed using computer vision software such as OpenCV™. Feature extraction can be based on the use of HoGs. HoGs can include feature descriptors and can be used to count occurrences of gradient orientation in localized portions or regions of the image. Other techniques can be used for counting occurrences of gradient orientation, including edge orientation histograms, scale-invariant feature transformation descriptors, etc. The AU recognition tasks can also be performed using Local Binary Patterns (LBP) and Local Gabor Binary Patterns (LGBP). The HoG descriptor represents the face as a distribution of intensity gradients and edge directions and is robust in its ability to translate and scale. Differing patterns, including groupings of cells of various sizes and arranged in variously sized cell blocks, can be used. For example, 4×4 cell blocks of 8×8 pixel cells with an overlap of half of the block can be used. Histograms of channels can be used, including nine channels or bins evenly spread over 0-180 degrees. In this example, the HoG descriptor on a 96×96 image is 25 blocks×16 cells×9 bins=3600, the latter quantity representing the dimension. AU occurrences can be rendered. The videos can be grouped into demographic datasets based on nationality and/or other demographic parameters for further detailed analysis. This grouping and other analyses can be facilitated via semiconductor-based logic.
A face 810 can be observed using a camera 830 in order to collect facial data that includes facial landmarks. The facial data can be collected from a plurality of people using one or more of a variety of cameras. As previously discussed, the camera or cameras can include a webcam, where a webcam can include a video camera, a still camera, a thermal imager, a CCD device, a smartphone camera, a three-dimensional camera, a depth camera, a light field camera, multiple webcams used to show different views of a person, or any other type of image capture apparatus that can allow captured data to be used in an electronic system. The quality and usefulness of the facial data that is captured can depend on the position of the camera 830 relative to the face 810, the number of cameras used, the illumination of the face, etc. In some cases, if the face 810 is poorly lit or over-exposed (e.g. in an area of bright light), the processing of the facial data to identify facial landmarks might be rendered more difficult. In another example, the camera 830 being positioned to the side of the person might prevent capture of the full face. Other artifacts can degrade the capture of facial data. For example, the person's hair, prosthetic devices (e.g. glasses, an eye patch, and eye coverings), jewelry, and clothing can partially or completely occlude or obscure the person's face. Data relating to various facial landmarks can include a variety of facial features. The facial features can comprise an eyebrow 820, an outer eye edge 822, a nose 824, a corner of a mouth 826, and so on. Multiple facial landmarks can be identified from the facial data that is captured. The facial landmarks that are identified can be analyzed to identify facial action units. The action units that can be identified can include AU02 outer brow raiser, AU14 dimpler, AU17 chin raiser, and so on. Multiple action units can be identified. The action units can be used alone and/or in combination to infer one or more mental states and emotions. A similar process can be applied to gesture analysis (e.g. hand gestures) with all of the analysis being accomplished or augmented by a mobile device, a server, semiconductor-based logic, and so on.
The facial regions that can be collected by the camera 930, sensor, or combination of cameras and/or sensors can include any of a variety of facial features. The facial features that can be included in the facial regions that are collected can include eyebrows 940, eyes 942, a nose 944, a mouth 946, ears, hair, texture, tone, and so on. Multiple facial features can be included in one or more facial regions. The number of facial features that can be included in the facial regions can depend on the desired amount of data to be captured, whether a face is in profile, whether the face is partially occluded or obstructed, etc. The facial regions that can include one or more facial features can be analyzed to determine facial expressions. The analysis of the facial regions can also include determining probabilities of occurrence of one or more facial expressions. The facial features that can be analyzed can also include textures, gradients, colors, shapes, etc. The facial features can be used to determine demographic data, where the demographic data can include age, ethnicity, culture, gender, etc. Multiple textures, gradients, colors, shapes, and so on, can be detected by the camera 930, sensor, or combination of cameras and sensors. Texture, brightness, and color, for example, can be used to detect boundaries in an image for detection of a face, facial features, facial landmarks, and so on.
A texture in a facial region can include facial characteristics, skin types, and so on. In some instances, a texture in a facial region can include smile lines, crow's feet, wrinkles, and so on. Another texture that can be used to evaluate a facial region can include a smooth portion of skin such as a smooth portion of a cheek. A gradient in a facial region can include values assigned to local skin texture, shading, etc. A gradient can be used to encode a texture, for example, by computing magnitudes in a local neighborhood or portion of an image. The computed values can be compared to discrimination levels, threshold values, and so on. The gradient can be used to determine gender, facial expression, etc. A color in a facial region can include eye color, skin color, hair color, and so on. A color can be used to determine demographic data, where the demographic data can include ethnicity, culture, age, gender, etc. A shape in a facial region can include shape of a face, eyes, nose, mouth, ears, and so on. As with color in a facial region, shape in a facial region can be used to determine demographic data including ethnicity, culture, age, gender, and so on.
The facial regions can be detected based on detection of edges, boundaries, and so on, of features that can be included in an image. The detection can be based on various types of analysis of the image. The features that can be included in the image can include one or more faces. A boundary can refer to a contour in an image plane where the contour can represent ownership of a particular picture element (pixel) from one object, feature, etc. in the image, to another object, feature, and so on, in the image. An edge can be a distinct, low-level change of one or more features in an image. That is, an edge can be detected based on a change, including an abrupt change, in color, brightness, etc. within an image. In embodiments, image classifiers are used for the analysis. The image classifiers can include algorithms, heuristics, and so on, and can be implemented using functions, classes, subroutines, code segments, etc. The classifiers can be used to detect facial regions, facial features, and so on. As discussed above, the classifiers can be used to detect textures, gradients, color, shapes, edges, etc. Any classifier can be used for the analysis, including, but not limited to, density estimation, support vector machines, logistic regression, classification trees, and so on. By way of example, consider facial features that can include the eyebrows 940. One or more classifiers can be used to analyze the facial regions that can include the eyebrows to determine a probability for either a presence or an absence of an eyebrow furrow. The probability can include a posterior probability, a conditional probability, and so on. The probabilities can be based on Bayesian Statistics or another statistical analysis technique. The presence of an eyebrow furrow can indicate that the person from whom the facial data can be collected is annoyed, confused, unhappy, and so on. In another example, consider facial features that can include a mouth 946. One or more classifiers can be used to analyze the facial region that can include the mouth to determine a probability for either a presence or an absence of mouth edges turned up to form a smile. Multiple classifiers can be used to determine one or more facial expressions.
The flow 1000, or portions thereof, can be implemented in semiconductor logic, accomplished using a mobile device, accomplished using a server device, and so on. The flow 1000 can be used to automatically detect a wide range of facial expressions. A facial expression can produce strong emotional signals that can indicate valence and discrete emotional states. The discrete emotional states can include contempt, doubt, defiance, happiness, fear, anxiety, and so on. The detection of facial expressions can be based on the location of facial landmarks. The detection of facial expressions can be based on determination of action units, where the action units are determined using FACS coding. The AUs can be used singly or in combination to identify facial expressions. Based on the facial landmarks, one or more AUs can be identified by number and intensity. For example, AU12 can be used to code a lip corner puller and can further be used to infer a smirk.
The flow 1000 begins by obtaining training image samples 1010. The image samples can include a plurality of images of one or more people. Human coders who are trained to correctly identify AU codes based on the FACS can code the images. The training or “known good” images can be used as a basis for training a machine learning technique. Once trained, the machine learning technique can be used to identify AUs in other images that can be collected using a camera, a sensor, and so on. The flow 1000 continues with receiving an image 1020. The image 1020 can be received from a camera, a sensor, and so on. As previously discussed, the camera or cameras can include a webcam, where a webcam can include a video camera, a still camera, a thermal imager, a CCD device, a phone camera, a three-dimensional camera, a depth camera, a light field camera, multiple webcams used to show different views of a person, or any other type of image capture apparatus that can allow captured data to be used in an electronic system. The image that is received can be manipulated in order to improve the processing of the image. For example, the image can be cropped, scaled, stretched, rotated, flipped, etc. in order to obtain a resulting image that can be analyzed more efficiently. Multiple versions of the same image can be analyzed. In some cases, the manipulated image and a flipped or mirrored version of the manipulated image can be analyzed alone and/or in combination to improve analysis. The flow 1000 continues with generating histograms 1030 for the training images and the one or more versions of the received image. The histograms can be based on a HoG or another histogram. As described in previous paragraphs, the HoG can include feature descriptors and can be computed for one or more regions of interest in the training images and the one or more received images. The regions of interest in the images can be located using facial landmark points, where the facial landmark points can include outer edges of nostrils, outer edges of the mouth, outer edges of eyes, etc. A HoG for a given region of interest can count occurrences of gradient orientation within a given section of a frame from a video.
The flow 1000 continues with applying classifiers 1040 to the histograms. The classifiers can be used to estimate probabilities, where the probabilities can correlate with an intensity of an AU or an expression. In some embodiments, the choice of classifiers used is based on the training of a supervised learning technique to identify facial expressions. The classifiers can be used to identify into which of a set of categories a given observation can be placed. The classifiers can be used to determine a probability that a given AU or expression is present in a given image or frame of a video. In various embodiments, the one or more AUs that are present include AU01 inner brow raiser, AU12 lip corner puller, AU38 nostril dilator, and so on. In practice, the presence or absence of multiple AUs can be determined. The flow 1000 continues with computing a frame score 1050. The score computed for an image, where the image can be a frame from a video, can be used to determine the presence of a facial expression in the image or video frame. The score can be based on one or more versions of the image 1020 or a manipulated image. The score can be based on a comparison of the manipulated image to a flipped or mirrored version of the manipulated image. The score can be used to predict a likelihood that one or more facial expressions are present in the image. The likelihood can be based on computing a difference between the outputs of a classifier used on the manipulated image and on the flipped or mirrored image, for example. The classifier that is used can be used to identify symmetrical facial expressions (e.g. smile), asymmetrical facial expressions (e.g. outer brow raiser), and so on.
The flow 1000 continues with plotting results 1060. The results that are plotted can include one or more scores for one or more frames computed over a given time t. For example, the plotted results can include classifier probability results from analysis of HoGs for a sequence of images and video frames. The plotted results can be matched with a template 1062. The template can be temporal and can be represented by a centered box function or another function. A best fit with one or more templates can be found by computing a minimum error. Other best-fit techniques can include polynomial curve fitting, geometric curve fitting, and so on. The flow 1000 continues with applying a label 1070. The label can be used to indicate that a particular facial expression has been detected in the one or more images or video frames which constitute the image that was received 1020. The label can be used to indicate that any of a range of facial expressions has been detected, including a smile, an asymmetric smile, a frown, and so on. Various steps in the flow 1000 may be changed in order, repeated, omitted, or the like without departing from the disclosed concepts. Various embodiments of the flow 1000 can be included in a computer program product embodied in a non-transitory computer readable medium that includes code executable by one or more processors. Various embodiments of the flow 1000, or portions thereof, can be included on a semiconductor chip and implemented in special purpose logic, programmable logic, and so on.
The flow 1100 includes obtaining videos containing faces 1110. The videos can be obtained using one or more cameras, where the cameras can include a webcam coupled to one or more devices employed by the one or more people using the web-based framework. The flow 1100 continues with extracting features from the individual responses 1120. The individual responses can include videos containing faces observed by the one or more webcams. The features that are extracted can include facial features such as an eyebrow, a nostril, an eye edge, a mouth edge, and so on. The feature extraction can be based on facial coding classifiers, where the facial coding classifiers output a probability that a specific facial action has been detected in a given video frame. The flow 1100 continues with performing unsupervised clustering of features 1130. The unsupervised clustering can be based on an event. The unsupervised clustering can be based on a K-Means, where the K of the K-Means can be computed using a Bayesian Information Criterion (BICk). It is possible, for example, to determine the smallest value of K that meets system requirements. Any other criterion for K can be used. The K-Means clustering technique can be used to group one or more events into various respective categories.
The flow 1100 includes characterizing cluster profiles 1140. The profiles can include a variety of facial expressions such as smiles, asymmetric smiles, eyebrow raisers, eyebrow lowerers, etc. The profiles can be related to a given event. For example, a humorous video can be displayed in the web-based framework and the video data of people who have opted-in can be collected. The characterization of the collected and analyzed video can depend in part on the number of smiles that occurred at various points throughout the humorous video. Similarly, the characterization can be performed on collected and analyzed videos of people viewing a news presentation. The characterized cluster profiles can be further analyzed based on demographic data. The number of smiles resulting from people viewing a humorous video can be compared to various demographic groups, where the groups can be formed based on geographic location, age, ethnicity, gender, and so on.
The flow 1100 can include determining mental state event temporal signatures. The mental state event temporal signatures can include information on rise time to facial expression intensity, fall time from facial expression intensity, duration of a facial expression, and so on. In some embodiments, the mental state event temporal signatures are associated with certain demographics, ethnicities, cultures, etc. The mental state event temporal signatures can be used to identify one or more of sadness, stress, happiness, anger, frustration, confusion, disappointment, hesitation, cognitive overload, focusing, engagement, attention, boredom, exploration, confidence, trust, delight, disgust, skepticism, doubt, satisfaction, excitement, laughter, calmness, curiosity, humor, depression, envy, sympathy, embarrassment, poignancy, or mirth. Various steps in the flow 1100 may be changed in order, repeated, omitted, or the like without departing from the disclosed concepts. Various embodiments of the flow 1100 can be included in a computer program product embodied in a non-transitory computer readable medium that includes code executable by one or more processors. Various embodiments of the flow 1100, or portions thereof, can be included on a semiconductor chip and implemented in special purpose logic, programmable logic, and so on.
Emotion analysis is a very complex task. Understanding and evaluating moods, emotions, or mental states requires a nuanced evaluation of facial expressions or other cues generated by people. Mental state analysis is important in many areas such as research, psychology, business, intelligence, law enforcement, and so on. The understanding of mental states can be used in a variety of fields, such as improving marketing analysis, assessing the effectiveness of customer service interactions and retail experiences, and evaluating the consumption of content such as movies and videos. Identifying points of frustration in a customer transaction can allow a company to take action to address the causes of the frustration. By streamlining processes, key performance areas such as customer satisfaction and customer transaction throughput can be improved, resulting in increased sales and revenues. In a content scenario, producing compelling content that achieves the desired effect (e.g. fear, shock, laughter, etc.) can result in increased ticket sales and/or increased advertising revenue. If a movie studio is producing a horror movie, it is desirable to know if the scary scenes in the movie are achieving the desired effect. By conducting tests in sample audiences, and analyzing faces in the audience, a computer-implemented method and system can process thousands of faces to assess the mental state at the time of the scary scenes. In many ways, such an analysis can be more effective than surveys that ask audience members questions, since audience members may consciously or subconsciously change answers based on peer pressure or other factors. However, spontaneous facial expressions can be more difficult to conceal. Thus, by analyzing facial expressions en masse in real time, important information regarding the mental state of the audience can be obtained.
Analysis of facial expressions is also a complex undertaking. Image data, where the image data can include facial data, can be analyzed to identify a range of facial expressions. The facial expressions can include a smile, frown, smirk, and so on. The image data and facial data can be processed to identify the facial expressions. The processing can include analysis of expression data, action units, gestures, mental states, physiological data, and so on. Facial data as contained in the raw video data can include information on one or more of action units, head gestures, smiles, brow furrows, squints, lowered eyebrows, raised eyebrows, attention, and the like. The action units can be used to identify smiles, frowns, and other facial indicators of expressions. Gestures can also be identified, and can include a head tilt to the side, a forward lean, a smile, a frown, as well as many other gestures. Other types of data including the physiological data can be obtained, where the physiological data can be obtained using a camera or other image capture device, without contacting the person or persons. Respiration, heart rate, heart rate variability, perspiration, temperature, and other physiological indicators of mental state can be determined by analyzing the images and video data.
Deep learning is a branch of machine learning which seeks to imitate in software the activity which takes place in layers of neurons in the neocortex of the human brain. This imitative activity can enable software to “learn” to recognize and identify patterns in data, where the data can include digital forms of images, sounds, and so on. The deep learning software is used to simulate the large array of neurons of the neocortex. This simulated neocortex, or artificial neural network, can be implemented using mathematical formulas that are evaluated on processors. With the ever-increasing capabilities of the processors, increasing numbers of layers of the artificial neural network can be processed.
Deep learning applications include processing of image data, audio data, and so on. Image data applications include image recognition, facial recognition, etc. Image data applications can include differentiating dogs from cats, identifying different human faces, and the like. The image data applications can include identifying moods, mental states, emotional states, and so on, from the facial expressions of the faces that are identified. Audio data applications can include analyzing audio such as ambient room sounds, physiological sounds such as breathing or coughing, noises made by an individual such as tapping and drumming, voices, and so on. The voice data applications can include analyzing a voice for timbre, prosody, vocal register, vocal resonance, pitch, loudness, speech rate, or language content. The voice data analysis can be used to determine one or more moods, mental states, emotional states, etc.
The artificial neural network which forms the basis for deep learning is based on layers. The layers can include an input layer, a convolution layer, a fully connected layer, a classification layer, and so on. The input layer can receive input data such as image data, where the image data can include a variety of formats including pixel formats. The input layer can then perform processing such as identifying boundaries of the face, identifying landmarks of the face, extracting features of the face, and/or rotating a face within the plurality of images. The convolution layer can represent an artificial neural network such as a convolutional neural network. A convolutional neural network can contain a plurality of hidden layers within it. A convolutional layer can reduce the amount of data feeding into a fully connected layer. The fully connected layer processes each pixel/data point from the convolutional layer. A last layer within the multiple layers can provide output indicative of mental state. The last layer of the convolutional neural network can be the final classification layer. The output of the final classification layer can be indicative of the mental states of faces within the images that are provided to the input layer.
Deep networks including deep convolutional neural networks can be used for facial expression parsing. A first layer of the deep network includes multiple nodes, where each node represents a neuron within a neural network. The first layer can receive data from an input layer. The output of the first layer can feed to a second layer, where the latter layer also includes multiple nodes. A weight can be used to adjust the output of the first layer which is being input to the second layer. Some layers in the convolutional neural network can be hidden layers. The output of the second layer can feed to a third layer. The third layer can also include multiple nodes. A weight can adjust the output of the second layer which is being input to the third layer. The third layer may be a hidden layer. Outputs of a given layer can be fed to the next layer. Weights adjust the output of one layer as it is fed to the next layer. When the final layer is reached, the output of the final layer can be a facial expression, a mental state, a characteristic of a voice, and so on. The facial expression can be identified using a hidden layer from the one or more hidden layers. The weights can be provided on inputs to the multiple layers to emphasize certain facial features within the face. The convolutional neural network can be trained to identify facial expressions, voice characteristics, etc. The training can include assigning weights to inputs on one or more layers within the multilayered analysis engine. One or more of the weights can be adjusted or updated during training. The assigning weights can be accomplished during a feed-forward pass through the multilayered neural network. In a feed-forward arrangement, the information moves forward, from the input nodes, through the hidden nodes and on to the output nodes. Additionally, the weights can be updated during a backpropagation process through the multilayered analysis engine.
Returning to the figure,
The network includes a collection of intermediate layers 1220. The multilayered analysis engine can include a convolutional neural network. Thus, the intermediate layers can include a convolution layer 1222. The convolution layer 1222 can include multiple sublayers, including hidden layers within it. The output of the convolution layer 1222 feeds into a pooling layer 1224. The pooling layer 1224 performs a data reduction, which makes the overall computation more efficient. Thus, the pooling layer reduces the spatial size of the image representation to reduce the number of parameters and computation in the network. In some embodiments, the pooling layer is implemented using filters of size 2×2, applied with a stride of two samples for every depth slice along both width and height, resulting in a reduction of 75-percent of the downstream node activations. The multilayered analysis engine can further include a max pooling layer 1224. Thus, in embodiments, the pooling layer is a max pooling layer, in which the output of the filters is based on a maximum of the inputs. For example, with a 2×2 filter, the output is based on a maximum value from the four input values. In other embodiments, the pooling layer is an average pooling layer or L2-norm pooling layer. Various other pooling schemes are possible.
The intermediate layers can include a Rectified Linear Units (RELU) layer 1226. The output of the pooling layer 1224 can be input to the RELU layer 1226. In embodiments, the RELU layer implements an activation function such as f(x)−max(0,x), thus providing an activation with a threshold at zero. In some embodiments, the RELU layer 1226 is a leaky RELU layer. In this case, instead of the activation function providing zero when x<0, a small negative slope is used, resulting in an activation function such as f(x)=1(x<0)(αx)+1(x>=0)(x). This can reduce the risk of “dying RELU” syndrome, where portions of the network can be “dead” with nodes/neurons that do not activate across the training dataset. The image analysis can comprise training a multilayered analysis engine using the plurality of images, wherein the multilayered analysis engine can include multiple layers that include one or more convolutional layers 1222 and one or more hidden layers, and wherein the multilayered analysis engine can be used for emotional analysis.
The system 1200 includes a fully connected layer 1230. The fully connected layer 1230 processes each pixel/data point from the output of the collection of intermediate layers 1220. The fully connected layer 1230 takes all neurons in the previous layer and connects them to every single neuron it has. The output of the fully connected layer 1230 provides input to a classification layer 1240. The output of the classification layer 1240 provides a facial expression and/or mental state as its output. Thus, a multilayered analysis engine such as the one depicted in
The cluster profiles 1302 can be generated based on the clusters that can be formed from unsupervised clustering, with time shown on the x-axis and intensity or frequency shown on the y-axis. The cluster profiles can be based on captured facial data, including facial expressions. The cluster profile 1320 can be based on the cluster 1310, the cluster profile 1322 can be based on the cluster 1312, and the cluster profile 1324 can be based on the cluster 1314. The cluster profiles 1320, 1322, and 1324 can be based on smiles, smirks, frowns, or any other facial expressions. The emotional states of the people who have opted-in to video collection can be inferred by analyzing the clustered facial expression data. The cluster profiles can be plotted with respect to time and can show a rate of onset, a duration, and an offset (rate of decay). Other time-related factors can be included in the cluster profiles. The cluster profiles can be correlated with demographic information, as described above.
The tags embedded in the webpage can be used for image analysis for emotional metric generation. The tags embedded in the website can also be used for speech analysis for cross-language mental state identification. Image analysis can include detection of facial expressions and can be performed for emotional metric generation. The facial expressions can be detected from people as they interact with the Internet. Image data, including facial images, is collected from a user interacting with a media presentation. Processors are used to analyze the image data and the media presentation to extract emotional content. Emotional intensity metrics are determined and retained in a storage component. The emotional intensity metrics are coalesced into a summary intensity metric, and the summary intensity metric is displayed on a screen. Once a tag is detected, a mobile device, a server, semiconductor-based logic, etc. can be used to evaluate associated facial expressions. A webpage 1400 can include a page body 1410, a page banner 1412, and so on. The page body can include one or more objects, where the objects can include text, images, videos, audio, and so on. The example page body 1410 shown includes a first image, image 1 1420; a second image, image 2 1422; a first content field, content field 1 1440; and a second content field, content field 2 1442. In practice, the page body 1410 can contain multiple images and content fields and can include one or more videos, one or more audio presentations, and so on. The page body can include embedded tags, such as tag 1 1430 and tag 2 1432. In the example shown, tag 1 1430 is embedded in image 1 1420, and tag 2 1432 is embedded in image 2 1422. In embodiments, multiple tags are embedded. Tags can also be embedded in content fields, in videos, in audio presentations, etc. When a user mouses over a tag or clicks on an object associated with a tag, the tag can be invoked. For example, when the user mouses over tag 1 1430, tag 1 1430 can then be invoked. Invoking tag 1 1430 can include enabling a camera coupled to a user's device to capture one or more images of the user as the user views a media presentation (or digital experience). In a similar manner, when the user mouses over tag 2 1432, tag 2 1432 can be invoked. Invoking tag 2 1432 can also include enabling the camera to capture images of the user. In other embodiments, other actions are taken based on invocation of the one or more tags. Invoking an embedded tag can initiate an analysis technique, post to social media, award the user a coupon or another prize, initiate mental state analysis, perform emotion analysis, and so on.
The invoking tags to collect images can be used for image analysis for emotional metric generation. The invoking tags to collect images can be used for people as they interact with various content provided to them, including content provided over the Internet. The tags can be related to analysis of mental state data for an individual. A mood dashboard can be displayed to the individual based on the analyzing. As previously stated, a media presentation can be a video, a webpage, and so on. A video 1402 can include one or more embedded tags, such as a tag 1460, another tag 1462, a third tag 1464, a fourth tag 1466, and so on. In practice, multiple tags can be included in the media presentation. The one or more tags can be invoked during the media presentation. The collection of the invoked tags can occur over time, as represented by a timeline 1450. When a tag is encountered in the media presentation, the tag can be invoked. When the tag 1460 is encountered, invoking the tag can enable a camera coupled to a user device to capture one or more images of the user viewing the media presentation. Invoking a tag can depend on opt-in by the user. For example, if a user has agreed to participate in a study by indicating an opt-in, then the camera coupled to the user's device can be enabled and one or more images of the user can be captured. If the user has not agreed to participate in the study and has not indicated an opt-in, then invoking the tag 1460 does not enable the camera nor capture images of the user during the media presentation. The user can indicate an opt-in for certain types of participation, where opting-in can be dependent on specific content in the media presentation. For example, the user could opt-in to participation in a study of political campaign messages and not opt-in for a particular advertisement study. In this case, tags that are related to political campaign messages, advertising messages, social media sharing, etc. and that enable the camera and image capture when invoked would be embedded in the media presentation social media sharing, and so on. However, tags embedded in the media presentation that are related to advertisements would not enable the camera when invoked. Various other situations of tag invocation are possible.
The capturing of images, videos, frames from video, etc. of one or more individuals results in substantial quantities of data that is stored for analysis, evaluation, comparison, aggregation, and other purposes. The image and video quality can vary depending on the capabilities of the machine or electronic device that is gathering the image and video data. The video frame rate can include 15 frames per second (fps), 30 fps, and so on. The data that is received by the one or more individuals, such as content provided by a content provider and delivered over the Internet from a website rendered for the one or more individuals, can also be stored. Further, key clicks, mouse clicks, tag invocations, and other directed and automatic user actions result in additional data. The result of the capturing of video data, content, user web journey information, and so on is that the volume of data increases over time.
The data, such as the video data collected from an individual, includes mental state data, facial data, and so on. The mental state data from the one or more individuals can be analyzed to determine one or more moods, one or more mental states, one or more emotional states, etc., for the one or more individuals. The purposes of the analysis can vary and can include determining whether a website, web content, and so on makes a given individual happy, sad, angry, and so on. Such analysis can compare recently collected data to data collected in the past, where the past can be earlier in the day, a previous day, an earlier week, last year, etc. This “data telescoping” can be useful to both the individual consumer of web content and to the content provider of the web and other content. The data telescoping can be used to recommend and/or direct an individual to a website that makes her or him happy, to avoid websites that induce anger, and so on. Additionally, the data telescoping can be used by a content provider to send to an individual content in which that individual is interested, to not send content that makes the individual angry, etc.
The value of the stored data changes over time. Current data can have the highest value and relevance, and can be stored in its entirety at a micro level. As the data ages, the typical trend is for the data to become less useful for such actions as predicting a current mental or emotional state in an individual, determining which content to provide, and so on. Various techniques can be used to determine what to do with the aging data. For example, after a week, the mental state data for an individual may be less relevant for determining current mental or emotional state, but can still maintain relevance for making comparisons of moods, emotions, mental states, determining trends, and so on. Over time, the data can be aggregated to time intervals. Such time intervals can include aggregating to one second intervals after a week, aggregating to the minute after a month, aggregating to an hour after a year, etc. The aggregation of data can be based on different techniques. One technique for data aggregation can include overall levels identified in the data such as whether the individual is happier, angrier, more confused, etc., when visiting a website or other content conduit. Another technique for data aggregation can include events such as numbers of smiles, eyebrow raises, scowls, etc. Aggregation of the data can also be based on filters used to identify data that should be kept, and other data that should be discarded.
The techniques used for the storage of the data are based on cost of storage, convenience of storage, uses of the data, and so on. Such data “warehousing” typically supports multiple uses of the data. A content provider may want to know the current mental and emotional states of an individual in order to provide that individual with content that will make that individual happy. The data storage accessed by the content provider would be fast and “nearby” for ready access, right now. By comparison, data scientists studying the collected data may be satisfied with slower, “farther away” storage. This latter class of storage provides for inexpensive storage of larger quantities of data than does the former class of storage.
The Internet 1510, intranet, or another wired, wireless, or hybrid computer network can be used for communication among the various devices and machines that comprise a system for speech analysis. A collecting device 1520 has a memory 1526 which stores instructions and one or more processors 1524 attached to the memory 1526, wherein the one or more processors 1524 can execute instructions. The collecting device 1520 can also have an internet connection to carry audio, utterances and mental states 1560, etc., and a display 1522 that can present various renderings and presentations to a user. The collecting device 1520 can collect utterances and mental state data from a plurality of people as they interact with a rendering. The collecting device 1520 can include a camera 1528. The camera 1528 can include a webcam, a video camera, a still camera, a thermal imager, a CCD device, a phone camera, a three-dimensional camera, a depth camera, a light field camera, multiple webcams used to show different views of a person, or any other type of image capture technique that can allow captured data to be used in an electronic system. In some embodiments, there are multiple collecting devices 1520 that each collect mental state data including utterances from one person or a plurality of people as they interact with a rendering. The collecting device 1520 can communicate with a training server 1530 and other machines over the internet 1510, some other computer network, or by another method suitable for communication between two computers. In some embodiments, the training machine 1530 functionality is embodied in the collecting device 1520.
The training machine 1530 can have an internet connection for individual training information 1562, a memory 1536 which stores instructions, and one or more processors 1534 attached to the memory 1536, wherein the one or more processors 1534 can execute instructions. The training machine 1530 can receive training information 1562 collected from one or more people as they produce utterances, interact with a rendering, etc., from the collecting device 1520 and can train a machine learning system using the first group of utterances and the associated first set of mental states. The machine learning system can include a support vector machine, artificial neural networks, convolutional neural networks (CNN), and so on. In some embodiments, the training machine 1530 also allows a user to view and evaluate the utterances, mental state information, training data, machine learning data, etc., that is associated with the rendering on a display 1532.
A storage device 1540 stores the first group of utterances and the associated first set of mental states, where the first group of utterances and the associated first set of mental states can include storage information 1564. The storage device can be connected to the Internet 1510 to exchange the storage information 1564. The storage device can include local storage, remote storage, distributed storage, cloud storage, and so on. The storage information can include the first group of utterances in a first language with an associated first set of mental states, the second group of utterances from a second language with an associated second set of mental states, a third group of utterances and an associated third group of utterances, and so on.
A processing machine 1550 can have a memory 1556 which stores instructions, and one or more processors 1554 attached to the memory 1556, wherein the one or more processors 1554 can execute instructions. The processing machine 1550 can use a connection to the Internet 1510, or another computer communication technique, to send and receive resulting information 1566. The processing machine 1550 can receive utterances and mental states information 1560, storage information 1564, training information 1562, etc. The processing machine can use a machine learning system that was trained to process a second group of utterances from a second language, wherein the processing determines a second set of mental states corresponding to the second group of utterances. The data and information can be rendered on a display 1552. The resulting information 1566 can include outputting the second set of mental states.
The system 1500 can include a computer program product embodied in a non-transitory computer readable medium for speech analysis, the computer program product comprising code which causes one or more processors to perform operations of: collecting, on a computing device, a first group of utterances in a first language with an associated first set of mental states; storing, on an electronic storage device, the first group of utterances and the associated first set of mental states; training a machine learning system using the first group of utterances and the associated first set of mental states that were stored; processing, on the machine learning system that was trained, a second group of utterances from a second language, wherein the processing determines a second set of mental states corresponding to the second group of utterances; and outputting the second set of mental states.
The system 1500 can include a computer system for speech analysis comprising: a memory which stores instructions; one or more processors attached to the memory wherein the one or more processors, when executing the instructions which are stored, are configured to: collect, on a computing device, a first group of utterances in a first language with an associated first set of mental states; store, on an electronic storage device, the first group of utterances and the associated first set of mental states; train a machine learning system using the first group of utterances and the associated first set of mental states that were stored; process, on the machine learning system that was trained, a second group of utterances from a second language, wherein the processing determines a second set of mental states corresponding to the second group of utterances; and output the second set of mental states.
Each of the above methods may be executed on one or more processors on one or more computer systems. Embodiments may include various forms of distributed computing, client/server computing, and cloud-based computing. Further, it will be understood that for each flow chart in this disclosure, the depicted steps or boxes are provided for purposes of illustration and explanation only. The steps may be modified, omitted, or re-ordered and other steps may be added without departing from the scope of this disclosure. Further, each step may contain one or more sub-steps. While the foregoing drawings and description set forth functional aspects of the disclosed systems, no particular arrangement of software and/or hardware for implementing these functional aspects should be inferred from these descriptions unless explicitly stated or otherwise clear from the context. All such arrangements of software and/or hardware are intended to fall within the scope of this disclosure.
The block diagrams and flowchart illustrations depict methods, apparatus, systems, and computer program products. Each element of the block diagrams and flowchart illustrations, as well as each respective combination of elements in the block diagrams and flowchart illustrations, illustrates a function, step or group of steps of the methods, apparatus, systems, computer program products and/or computer-implemented methods. Any and all such functions may be implemented by computer program instructions, by special-purpose hardware-based computer systems, by combinations of special purpose hardware and computer instructions, by combinations of general purpose hardware and computer instructions, and so on. Any and all of which may be generally referred to herein as a “circuit,” “module,” or “system.”
A programmable apparatus which executes any of the above-mentioned computer program products or computer implemented methods may include one or more microprocessors, microcontrollers, embedded microcontrollers, programmable digital signal processors, programmable devices, programmable gate arrays, programmable array logic, memory devices, application specific integrated circuits, or the like. Each may be suitably employed or configured to process computer program instructions, execute computer logic, store computer data, and so on.
It will be understood that a computer may include a computer program product from a computer-readable storage medium and that this medium may be internal or external, removable and replaceable, or fixed. In addition, a computer may include a Basic Input/Output System (BIOS), firmware, an operating system, a database, or the like that may include, interface with, or support the software and hardware described herein.
Embodiments of the present invention are not limited to applications involving conventional computer programs or programmable apparatus that run them. It is contemplated, for example, that embodiments of the presently claimed invention could include an optical computer, quantum computer, analog computer, or the like. A computer program may be loaded onto a computer to produce a particular machine that may perform any and all of the depicted functions. This particular machine provides a means for carrying out any and all of the depicted functions.
Any combination of one or more computer readable media may be utilized. The computer readable medium may be a non-transitory computer readable medium for storage. A computer readable storage medium may be electronic, magnetic, optical, electromagnetic, infrared, semiconductor, or any suitable combination of the foregoing. Further computer readable storage medium examples may include an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random-access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM, Flash, MRAM, FeRAM, or phase change memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing. In the context of this document, a computer readable storage medium may be any tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device.
It will be appreciated that computer program instructions may include computer executable code. A variety of languages for expressing computer program instructions may include without limitation C, C++, Java, JavaScript™, ActionScript™, assembly language, Lisp, Perl, Tcl, Python, Ruby, hardware description languages, database programming languages, functional programming languages, imperative programming languages, and so on. In embodiments, computer program instructions may be stored, compiled, or interpreted to run on a computer, a programmable data processing apparatus, a heterogeneous combination of processors or processor architectures, and so on. Without limitation, embodiments of the present invention may take the form of web-based computer software, which includes client/server software, software-as-a-service, peer-to-peer software, or the like.
In embodiments, a computer may enable execution of computer program instructions including multiple programs or threads. The multiple programs or threads may be processed more or less simultaneously to enhance utilization of the processor and to facilitate substantially simultaneous functions. By way of implementation, any and all methods, program codes, program instructions, and the like described herein may be implemented in one or more thread. Each thread may spawn other threads, which may themselves have priorities associated with them. In some embodiments, a computer may process these threads based on priority or other order.
Unless explicitly stated or otherwise clear from the context, the verbs “execute” and “process” may be used interchangeably to indicate execute, process, interpret, compile, assemble, link, load, or a combination of the foregoing. Therefore, embodiments that execute or process computer program instructions, computer-executable code, or the like may act upon the instructions or code in any and all of the ways described. Further, the method steps shown are intended to include any suitable method of causing one or more parties or entities to perform the steps. The parties performing a step, or portion of a step, need not be located within a particular geographic location or country boundary. For instance, if an entity located within the United States causes a method step, or portion thereof, to be performed outside of the United States then the method is considered to be performed in the United States by virtue of the entity causing the step to be performed.
While the invention has been disclosed in connection with preferred embodiments shown and described in detail, various modifications and improvements thereon will become apparent to those skilled in the art. Accordingly, the spirit and scope of the present invention is not to be limited by the foregoing examples, but is to be understood in the broadest sense allowable by law.
Claims
1. A computer-implemented method for speech analysis comprising:
- collecting, on a computing device, a first group of utterances in a first language with an associated first set of mental states;
- storing, on an electronic storage device, the first group of utterances and the associated first set of mental states;
- training a machine learning system using the first group of utterances and the associated first set of mental states that were stored;
- processing, on the machine learning system that was trained, a second group of utterances from a second language, wherein the processing determines a second set of mental states corresponding to the second group of utterances; and
- outputting the second set of mental states.
2. The method of claim 1 further comprising learning, from the first group of utterances and associated first set of mental states and the second group of utterances and associated second set of mental states, to facilitate determining an associated third set of mental states from a third group of utterances.
3. The method of claim 1 further comprising outputting a series of heuristics, based on correspondence between the first group of utterances and the associated first set of mental states.
4. The method of claim 1 wherein the machine learning system includes a deep learning system.
5. The method of claim 4 wherein the machine learning system performs convolving.
6. The method of claim 1 wherein the machine learning system includes a convolutional neural network.
7. The method of claim 1 wherein the first language and the second language are substantially similar.
8. The method of claim 7 wherein the first language and the second language are identical.
9. The method of claim 1 wherein the first language and the second language are different.
10. The method of claim 1 further comprising refining the training of the machine learning system based on one or more additional groups of utterances in the first language or the second language.
11. The method of claim 1 wherein the training comprises segmenting silence from speech in the second group of utterances.
12. The method of claim 1 wherein the training comprises extracting low-level acoustic descriptors from short, overlapping speech segments from the second group of utterances.
13. The method of claim 12 further comprising applying statistical functions to resolve low-level acoustic descriptors over longer speech segments.
14. The method of claim 12 further comprising extracting contextual information from neighboring speech segments.
15. The method of claim 12 further comprising feeding extracted features to a classifier for determining mental states.
16. The method of claim 1 wherein the training comprises estimating mental state metrics over successive, overlapped speech segments.
17. The method of claim 16 further comprising fusing the mental state metrics that were estimated to produce a smoothed mental state metric.
18. The method of claim 16 wherein the successive, overlapped speech segments are windowed around 1200 ms.
19. The method of claim 1 wherein the first group of utterances includes non-speech vocalizations.
20-21. (canceled)
22. The method of claim 1 wherein the outputting is used for developing cross-cultural conversational agents.
23. The method of claim 22 wherein the cross-cultural conversational agents are used in vehicular control.
24. The method of claim 1 further comprising training an application for use with a third language distinct from the first language and the second language.
25. The method of claim 1 further comprising developing cross-linguistic models based on the outputting.
26. The method of claim 25 further comprising training the cross-linguistic models based on one or more human reactions to an application using the cross-linguistic models.
27. (canceled)
28. A computer program product embodied in a non-transitory computer readable medium for speech analysis, the computer program product comprising code which causes one or more processors to perform operations of:
- collecting, on a computing device, a first group of utterances in a first language with an associated first set of mental states;
- storing, on an electronic storage device, the first group of utterances and the associated first set of mental states;
- training a machine learning system using the first group of utterances and the associated first set of mental states that were stored;
- processing, on the machine learning system that was trained, a second group of utterances from a second language, wherein the processing determines a second set of mental states corresponding to the second group of utterances; and
- outputting the second set of mental states.
29. A computer system for speech analysis comprising:
- a memory which stores instructions;
- one or more processors attached to the memory wherein the one or more processors, when executing the instructions which are stored, are configured to: collect, on a computing device, a first group of utterances in a first language with an associated first set of mental states; store, on an electronic storage device, the first group of utterances and the associated first set of mental states; train a machine learning system using the first group of utterances and the associated first set of mental states that were stored; process, on the machine learning system that was trained, a second group of utterances from a second language, wherein the processing determines a second set of mental states corresponding to the second group of utterances; and output the second set of mental states.
Type: Application
Filed: Nov 30, 2018
Publication Date: Jun 6, 2019
Applicant: Affectiva, Inc. (Boston, MA)
Inventors: Taniya Mishra (New York, NY), Islam Faisal (Kafr El-Zayat), Mohamed Ezzeldin Abdelmonem Ahmed Mohamed (Cairo)
Application Number: 16/206,135
