Abstract: Embodiments provide methods and apparatus for improving responses of automated conversational agents. The method includes generating a vector representation of a conversational input provided by a user. The vector representation is used to determine an intent of the conversational input. Further, annotators generate bait sentences that cover multiple aspects of the intent. Then, sentences in a data pool are accessed. The bait sentences and the data pool sentences are converted into a first and a second set of vector representations, respectively. The first and the second set of vector representations are compared to retrieve a list of similar sentences. The list of similar sentences includes one or more sentences of the data pool that are semantically similar to the bait sentences. The list of similar sentences is analyzed for updating the intent data and thereby improve the responses.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for connecting an interactive wearable device with a network. In one aspect, a method includes loading content from a playlist; recognizing contextual information relating to the content; determining the location of the user; requesting supplemental content via a network based on the contextual information and the location; displaying supplemental information to a user; interacting with the supplemental information at least in part via an interactive apparatus.

Abstract: A method of creating content adapted multimedia guidance for a web document, comprising: analyzing location and type of dynamic components in a web document for the creation of a flow dataset, that maps at least one trigger node representing a user initiated trigger event and at least one action node linked to the trigger node and representing instruction to present guidance related to one of the dynamic components in response to the user initiated trigger event; identifying a loading event during which the web document is loaded on a client terminal of a user; extracting at least one generated value from the dynamic components; building an audio sequence associated with the action node by selecting audio segments according to the generated value and arranging the audio segments according to the flow dataset; and when the trigger is activated, playing the audio sequence on the client terminal.

Abstract: In some implementations, a user device can be configured to create media messages with automatic titling. For example, a user can create a media messaging project that includes multiple video clips. The video clips can be generated based on video data and/or audio data captured by the user device and/or based on pre-recorded video data and/or audio data obtained from various storage locations. When the user device captures the audio data for a clip, the user device can obtain a speech-to-text transcription of the audio data in near real time and present the transcription data (e.g., text) overlaid on the video data while the video data is being captured or presented by the user device.

Abstract: Embodiments relate to a headset that filters sounds according to a direction of a gaze of a user wearing the headset. The user wears the headset including an eye tracking unit and one or more microphones. The eye tracking unit tracks an orientation of an eye of the user to determine the direction of the gaze of the user. The direction of the gaze may be different from a facing direction of the headset. According to the determined direction of the gaze of the user, input sound signals generated by the microphones can be beamformed to amplify or emphasize sound originating from the direction of the gaze.

Abstract: A method and an apparatus for protecting an application program password of a mobile terminal are described. The method includes: positioning an application interface view in an application program of a mobile terminal, so as to obtain a currently activated application interface view; determining whether the currently activated application interface view is an input box; and if yes, acquiring a key input event triggered in the input box and obtaining an input character by using the key input event; and replacing the character with a fake password character. A related apparatus is also described.

Abstract: According to an embodiment, a verification system includes a storage controller, first and second receivers, a comparator, a response constructor, a response generator, and an output controller. The storage controller stores, in a storage, first response data and first situation data associated with the first response data. The first receiver receives second response data. The comparator determines a similarity between second situation data indicating a second context for using the second response data and the first situation data. The response constructor constructs response content information comprising the second response data and the first response data associated with the first situation data having the similarity equal to or greater than a threshold. The second receiver receives speech data. The response generator generates a response sentence corresponding to the speech data using the response content information. The output controller outputs for display one or more response sentences.

Abstract: A device includes a memory, a receiver, a processor, and a display. The memory is configured to store a speaker model. The receiver is configured to receive an input audio signal. The processor is configured to determine a first confidence level associated with a first portion of the input audio signal based on the speaker model. The processor is also configured to determine a second confidence level associated with a second portion of the input audio signal based on the speaker model. The display is configured to present a graphical user interface associated with the first confidence level or associated with the second confidence level.

Abstract: A computer-implemented method includes associating, using a processor, one or more words in an electronic agenda template to at least one agenda item indicative of a point for discussion. The processor captures a real-time interaction comprising speech from one or more participants of a plurality of discussion participants into a digital representation. The processor isolates a portion of the real-time interaction from the digital representation. The portion is associated with a single speaker of the plurality of discussion participants. The processor makes at least one match between an isolated portion of the real-time interaction and the at least one agenda item. The processor determines an intent of the single speaker from the isolated portion and matching the determined intent of the single speaker to the at least one agenda item on the electronic agenda template, and generates discussion minutes output based on the matched intent and agenda item.

Abstract: A method, performed by at least one computer, the method comprising using the at least one computer to perform acts of accessing information specifying at least one user-specified condition specified by a user and at least one corresponding user-specified action, the user-specified action to be performed when the user-specified condition is met; determining whether the at least one user-specified condition is met; and when it is determined that the at least one user-specified condition is met, causing a virtual assistant executing on a mobile device different from the at least one computer to perform the at least one user-specified action.

Abstract: A method includes determining a content of a speech of a user on the basis of text data including a word string acquired as a recognition result of the speech, selecting a correction method suitable for the determined content of the speech from among multiple correction methods for correcting false recognition contained in the text data, and correcting the false recognition contained in the text data using the selected correction method.

Abstract: A mechanism is provided for delivering troubleshooting support that captures a user's interactions with a local computing device. Responsive to receiving a set of events associated with an issue being experienced by the user of the local computing device, each event in the set of events is presented to a technician. Each event in the set of events is flagged as either correct or incorrect. Responsive to the incorrect flagged event being an operation performed by the local computing device, a communicative connecting is established to the local computing device in order that the technician corrects a future operation corresponding to the operation performed. Responsive to the incorrect flagged event being an action performed by the user of the local computing device, an analysis is sent in order to inform the user of incorrect actions so as to not perform the incorrect actions in the future.

Abstract: Symbol sequences are estimated using a computer-implemented method including detecting one or more candidates of a target symbol sequence from a speech-to-text data, extracting a related portion of each candidate from the speech-to-text data, detecting repetition of at least a partial sequence of each candidate within the related portion of the corresponding candidate, labeling the detected repetition with a repetition indication, and estimating whether each candidate is the target symbol sequence, using the corresponding related portion including the repetition indication of each of the candidates.

Abstract: A method to determine whether a user is experiencing difficulty visually resolving content is disclosed. The method includes capturing one or more images of the user while the user is viewing the content. The method also includes obtaining facial parameters related to a visual acuity of the user from the captured one or more images. The method further includes determining whether the user is experiencing difficulty visually resolving the content based on the obtained one or more facial parameters. The method is implemented in a device such as a smartphone, tablet computer, or television. The facial parameters include information about the extent to which the user has their eyes open or closed, whether the user is wearing glasses, and the distance at which the user is viewing the content.

Abstract: In one aspect, a method includes (a) accessing, by a computing system, a social media (SM) platform, (b) selecting a SM content item published on the SM platform, (c) responsive to the selection, accessing a webpage of the SM platform, (d) detecting that the webpage contains a nested link object, (e) selecting the nested link object to display the hidden SM content item(s) on the webpage, (f) responsive to selecting the nested link object, evaluating the displayed hidden SM content item(s) on the webpage to determine whether the selected SM content item is displayed on the webpage, (g) repeating acts (d)-(f) until the computing system determines that the selected SM content item is displayed on the webpage, and (h) responsive to determining that the selected SM content item is displayed on the webpage, retrieving the selected SM content item via the webpage.

Abstract: The present technology automatically shares materials at the start of a videoconference without requiring a participant to find the materials or instruct the videoconferencing application to share the materials. The conference materials can be automatically shared without any conference participant involvement. The present technology automatically associates materials included in a calendar invitation to the conference or in a shared space referenced in the calendar invitation. These materials can be automatically shared when the conference launches.

Abstract: Methods, computer program products, and systems are presented. The method computer program products, and systems can include, for instance: obtaining an audio input, the audio input representing sound emanating from a key press based user interface; generating a context pattern based on the audio input; performing classification of the context pattern to classify the context pattern as belonging to a signature pattern classification, wherein the signature pattern classification specifies a user activity; and providing an output based on the performing classification.

Abstract: Techniques for using a dynamic wakeword detection threshold are described. A server(s) may receive audio data corresponding to an utterance from a device in response to the device detecting a wakeword using a wakeword detection threshold. The server(s) may then determine the device should use a lower wakeword detection threshold for a duration of time. In addition to sending the device output data responsive to the utterance, the server(s) may send the device an instruction to use the lower wakeword detection threshold for the duration of time. Alternatively, the server(s) may train a machine learning model to determine when the device should use a lower wakeword detection threshold. The server(s) may send the trained machine learned model to the device for use at runtime.

Abstract: The present disclosure generally relates to dynamic phrase expansion for language input. In one example process, a user input comprising text of a first symbolic system is received. The process determines, based on the text, a plurality of sets of one or more candidate words of a second symbolic system. The process determines, based on at least a portion of the plurality of sets of one or more candidate words, a plurality of expanded candidate phrases. Each expanded candidate phrase comprises at least one word of a respective set of one or more candidate words of the plurality of sets of one or more candidate words and one or more predicted words based on the at least one word of the respective set of one or more candidate words. One or more expanded candidate phrases of the plurality of expanded candidate phrases are displayed for user selection.

Abstract: Disclosed techniques provide for the management of immersion levels and notifications of conference activities that are of interest to a user. In some embodiments, a system can monitor a conference session, that a user is not attending, on the user's behalf to determine when activity that is of interest to the user occurs during the conference session. The system can then send a notification to the user to inform her of the identified activity. In some embodiments, the notification may enable the user to control her immersion level with respect to the conference session. For example, the notification may include user interface element(s) that the user may select to observe or even join the conference session. In some embodiments, the notification may enable the user to directly share content with attendees of the conference session without having to observe or join the conference session.

Abstract: Method and systems are provided for interpreting speech data. A method and system for recognizing speech involving a filter module to generate a set of processed audio data based on raw audio data; a translation module to provide a set of translation results for the raw audio data; and a decision module to select the text data that represents the raw audio data. A method for minimizing noise in audio signals received by a microphone array is also described. A method and system of automatic entry of data into one or more data fields involving receiving a processed audio data; and operating a processing module to: search in a trigger dictionary for a field identifier that corresponds to the trigger identifier; identify a data field associated with a data field identifier corresponding to the field identifier; and providing content data associated with the trigger identifier to the identified data field.

Abstract: Technology of the disclosure may facilitate user discovery of various voice-based action queries that can be spoken to initiate computer-based actions, such as voice-based action queries that can be provided as spoken input to a computing device to initiate computer-based actions that are particularized to content being viewed or otherwise consumed by the user on the computing device. Some implementations are generally directed to determining, in view of content recently viewed by a user on a computing device, at least one suggested voice-based action query for presentation via the computing device. Some implementations are additionally or alternatively generally directed to receiving at least one suggested voice-based action query at a computing device and providing the suggested voice-based action query as a suggestion in response to input to initiate providing of a voice-based query via the computing device.

Abstract: A method to transcribe communications may include obtaining audio data originating at a first device during a communication session between the first device and a second device and providing the audio data to an automated speech recognition system configured to transcribe the audio data. The method may further include obtaining multiple hypothesis transcriptions generated by the automated speech recognition system. Each of the multiple hypothesis transcriptions may include one or more words determined by the automated speech recognition system to be a transcription of a portion of the audio data. The method may further include determining one or more consistent words that are included in two or more of the multiple hypothesis transcriptions and in response to determining the one or more consistent words, providing the one or more consistent words to the second device for presentation of the one or more consistent words by the second device.

Abstract: Systems and methods of validating transcriptions of natural language content using crowdsourced validation jobs are provided herein. In various implementations, a transcription pair comprising natural language content and text corresponding to a transcription of the natural language content may be gathered. A group of validation devices may be selected for reviewing the transcription pair. A crowdsourced validation job may be created for the group of validation devices. The crowdsourced validation job may be provided to the group of validation devices. One or more votes representing whether or not the text accurately represents the natural language content may be received from the group of validation devices. Based on the one or more votes received, the transcription pair may be stored in a validated transcription library, which may be used to process end-user voice data.

Abstract: Methods, computer program products, and systems are presented. The method computer program products, and systems can include, for instance: obtaining an audio input, the audio input representing sound emanating from a key press based user interface; generating a context pattern based on the audio input; performing classification of the context pattern to classify the context pattern as belonging to a signature pattern classification, wherein the signature pattern classification specifies a user activity; and providing an output based on the performing classification.

Abstract: In some aspects, a method of using a virtual medical assistant to assist a medical professional, the virtual medical assistant implemented, at least in part, by at least one processor of a host device capable of connecting to at least one network is provided. The method comprises receiving free-form instruction from the medical professional, providing the free-form instruction for processing to assist in identifying from the free-form instruction at least one medical task to be performed, obtaining identification of at least one impediment to performing the at least one medical task, and inferring at least some information needed to overcome the at least one impediment.

Abstract: Systems and processes are disclosed for handling a multi-part voice command for a virtual assistant. Speech input can be received from a user that includes multiple actionable commands within a single utterance. A text string can be generated from the speech input using a speech transcription process. The text string can be parsed into multiple candidate substrings based on domain keywords, imperative verbs, predetermined substring lengths, or the like. For each candidate substring, a probability can be determined indicating whether the candidate substring corresponds to an actionable command. Such probabilities can be determined based on semantic coherence, similarity to user request templates, querying services to determine manageability, or the like. If the probabilities exceed a threshold, the user intent of each substring can be determined, processes associated with the user intents can be executed, and an acknowledgment can be provided to the user.

Abstract: One or more embodiments of techniques or systems for privacy management are provided herein. In one or more embodiments, privacy management may be provided by adjusting the way a system provides notifications, handles communications, or executes applications based on the presence or positioning of one or more occupants within a vehicle. In other words, a vehicle may provide notifications in a first manner when merely a driver is present but provide notifications in a second manner (e.g., different than the first manner) when an additional occupant is detected. Sensors may be utilized to detect the presence or positioning of occupants and communications or applications may be managed accordingly. In this manner, privacy management is provided, thus enhancing security of applications or communications utilized by occupants of a vehicle, such as a driver, for example.

Abstract: Systems, methods, and devices related to establishing voice identity profiles for use with voice-controlled devices are provided. The embodiments disclosed enhance user experience by customizing the enrollment process to utilize voice recognition for each user based on historical information which can be used in the selection process of phrases a user speaks during enrollment of a voice recognition function or skill. The selection process can utilize phrases that have already been spoken to the electronic device; it can utilize phrases, contacts, or other personalized information it can obtain from the user account of the person enrolling; it can use any of the information just described to select specific words to enhance the probably of achieving higher phonetic matches based on words the individual user is more likely to speak to the device.

Abstract: A voice to text model used by a voice-enabled electronic device is dynamically and in a context-sensitive manner updated to facilitate recognition of entities that potentially may be spoken by a user in a voice input directed to the voice-enabled electronic device. The dynamic update to the voice to text model may be performed, for example, based upon processing of a first portion of a voice input, e.g., based upon detection of a particular type of voice action, and may be targeted to facilitate the recognition of entities that may occur in a later portion of the same voice input, e.g., entities that are particularly relevant to one or more parameters associated with a detected type of voice action.

Abstract: Coordinated operation of a voice-controlled device and an accessory device in an environment is described. A remote system processes audio data it receives from the voice-controlled device in the environment to identify a first intent associated with a first domain, a second intent associated with a second domain, and a named entity associated with the audio data. The remote system sends, to the voice-controlled device, first information for accessing main content associated with the named entity, and a first instruction corresponding to the first intent. The remote system also sends, to the accessory device, second information for accessing control information or supplemental content associated with the main content, and a second instruction corresponding to the second intent. The first and second instructions, when processed by the devices in the environment, cause coordinated operation of the voice-controlled device and the accessory device.

Abstract: Disclosed herein is a system for compensating for dialects and accents comprising an automatic speech recognition system comprising an automatic speech recognition device that is operative to receive an utterance in an acoustic format from a user with a user interface; a speech to text conversion engine that is operative to receive the utterance from the automatic speech recognition device and to prepare a textual statement of the utterance; and a correction database that is operative to store textual statements of all utterances; where the correction database is operative to secure a corrected transcript of the textual statement of the utterance from the speech to text conversion engine and adds it to the corrections database if the corrected transcript of the textual statement of the utterance is not available.

Abstract: A method for natural language interaction includes recording speech provided by a human user. The recorded speech is translated into a machine-readable natural language input relating to an interaction topic. An interaction timer is maintained that tracks a length of time since a last machine-readable natural language input referring to the interaction topic was translated. Based on a current value of the interaction timer being greater than an interaction engagement threshold, a message relating to the interaction topic is delivered with a first natural language phrasing that includes an interaction topic reminder. Based on the current value of the interaction timer being less than the interaction engagement threshold, the message relating to the interaction topic is delivered with a second natural language phrasing that lacks the interaction topic reminder.

Abstract: Techniques described herein include a communication and device discovery platform though which Internet of Things (IoT) devices may discover other IoT devices, ascertain device statuses, purposes, and the functions and operations supported by other IoT devices, and collaborate with other IoT devices to request and initiate particular functions and operations. Using the platform and corresponding functionality described herein, an IoT device may announce to another nearby IoT devices its general description, availability, required permissions, as well as a listing and description of functions it may perform. Other authorized IoT devices may make use of the broadcasted functions, using artificial intelligence, natural language processing, and additional techniques that enable independent and decentralized IoT devices to discover the purposes, supported functions and operations, and invoke the functionality of other IoT devices.

Abstract: Systems and processes for identifying of a voice input providing one or more user credentials are provided. In one example process, a voice input can be received. A first character, a phrase identifying a second character, and a word can be identified based on the voice input. In response to the identification, the first character, the second character, and the word can be converted to text. The text can be caused to display, with a display, in a sequence corresponding to an order of the first character, the second character, and the word in the voice input.

Abstract: Media content can be created and/or modified using a network-accessible platform. Scripts for content-based experiences could be readily created using one or more interfaces generated by the network-accessible platform. For example, a script for a content-based experience could be created using an interface that permits triggers to be inserted directly into the script. Interface(s) may also allow different media formats to be easily aligned for post-processing. For example, a transcript and an audio file may be dynamically aligned so that the network-accessible platform can globally reflect changes made to either item. User feedback may also be presented directly on the interface(s) so that modifications can be made based on actual user experiences.

Abstract: Techniques and systems for synchronized accessibility for client devices in an online conference are described. For example, a conferencing system receives presentation content and audio content as part of the online conference from a client device. The conferencing system generates sign language content by converting audio in the audio content to sign language. The conferencing system then synchronizes display of the sign language content with the presentation content in a user interface based on differences in durations of segments of the audio content from durations of corresponding segments of the sign language content. Then, the conferencing system outputs the sign language content as synchronized with the presentation content, such as to a viewer client device that requested the sign language content, or to storage for later access by viewers that request sign language content.

Abstract: A system and method for allowing user intervention in a speech recognition pipeline is presented. Embodiments may include receiving, at a computing device, a speech signal at a speech recognition engine, the speech signal being associated with an application. Embodiments may further include generating one or more suggested speech results at the speech recognition engine, the suggested speech results based upon, at least in part, the speech signal. Embodiments may also include displaying, at a graphical user interface associated with the computing device, the one or more suggested speech results prior to applying a final speech result. Embodiments may further include receiving a non voice-based selection of at least one of the one or more suggested speech results and applying the non voice-based selection to the application.

Abstract: In some example embodiments, a graphical user interface (GUI) is caused to be displayed on a computing device of a user. The GUI can be configured to enable the user to submit an identification of a dataset and at least one configuration parameter. The identification of the data source and the at least one configuration parameter can be received via the GUI on the computing device. A sampling algorithm can be configured based on the at least one configuration parameter. A sample of data from the dataset can be generated using the configured sampling algorithm. The time parameter can limit an execution time of the generating of the sample of data to within a maximum amount of time indicated by the user.

Abstract: A parking system includes at least one first, near-field broadcasting device emitting a first radio field and at least one second, near-field broadcasting device emitting a second radio field. The system operates at least one access device. The first near-field broadcasting device transmits a first access data record, which is a function of a second access data record transmitted by the second near-field broadcasting device. An evaluation device identifies an associated access device in the system based on an evaluation of the functional relationship between the first access data record and the second access data record.

Abstract: An input support apparatus of an embodiment includes a template storage unit configured to store a form template that is a template for form data having one or more slots to which item values are input in correspondence with item names, the form template describing item names of the respective slots and alternatives of an alternative type slot in which an item value is selected from a plurality of alternatives together with respective readings thereof; an acquisition unit configured to acquire recognition result data obtained by speech recognition performed on utterance of a user, the recognition result data containing a transcription and a reading; and a determination unit configured to determine the item values to be input to the slots of the form data based on the reading of the recognition result data and the readings of the item names and the alternatives described in the form template.

Abstract: Disclosed is a glass-type terminal including: a frame portion that is configured to be wearable on the head of a user; a microphone that is combined with the frame portion and into which voice contents are input; a storage unit that is combined with the frame portion and in which at least one piece of schedule information relating to the voice contents are stored; and a controller that, when the voice contents are input into the microphone, analyzes the voice contents being input and extracts the schedule information associated with the analyzed voice contents from the storage unit. Then, in the glass-type terminal, the controller may determine whether or not the extracted schedule information is output, depending on whether or not the extracted schedule information and voice information included in the voice contents being input agree with each other.

Abstract: Assisting input from a keyboard is described. In an embodiment, a processor receives a plurality of key-presses from the keyboard comprising alphanumeric data for input to application software executed at the processor. The processor analyzes the plurality of key-presses to detect at least one predefined typing pattern, and, in response, controls a display device to display a representation of at least a portion of the keyboard in association with a user interface of the application software. In another embodiment, a computer device has a keyboard and at least one sensor arranged to monitor at least a subset of keys on the keyboard, and detect an object within a predefined distance of a selected key prior to activation of the selected key. The processor then controls the display device to display a representation of a portion of the keyboard comprising the selected key.

Abstract: Disclosed herein are systems, methods, and non-transitory computer-readable storage media for processing speech. A system configured to practice the method monitors user utterances to generate a conversation context. Then the system receives a current user utterance independent of non-natural language input intended to trigger speech processing. The system compares the current user utterance to the conversation context to generate a context similarity score, and if the context similarity score is above a threshold, incorporates the current user utterance into the conversation context. If the context similarity score is below the threshold, the system discards the current user utterance. The system can compare the current user utterance to the conversation context based on an n-gram distribution, a perplexity score, and a perplexity threshold. Alternately, the system can use a task model to compare the current user utterance to the conversation context.

Abstract: Provided are systems, computer-implemented methods, and computer-program products for a multi-lingual device, capable of receiving verbal input in multiple languages, and further capable of providing conversational responses in multiple languages. In various implementations, the multi-lingual device includes an automatic speech recognition engine capable of receiving verbal input in a first natural language and providing a textual representation of the input and a confidence value for the recognition. The multi-lingual device can also include a machine translation engine, capable of translating textual input from the first natural language into a second natural language. The machine translation engine can output a confidence value for the translation. The multi-lingual device can further include natural language processing, capable of translating from the second natural language to a computer-based language.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for processing audio waveforms. In some implementations, a time-frequency feature representation is generated based on audio data. The time-frequency feature representation is input to an acoustic model comprising a trained artificial neural network. The trained artificial neural network comprising a frequency convolution layer, a memory layer, and one or more hidden layers. An output that is based on output of the trained artificial neural network is received. A transcription is provided, where the transcription is determined based on the output of the acoustic model.

Abstract: A service providing apparatus including an occupant detector configured to detect presence of each of a plurality of occupants in a vehicle and a control unit including a CPU and a memory coupled to the memory, wherein the CPU and the memory are configured to perform: estimating an individual feeling of the each of the plurality of occupants detected by the occupant detector; estimating a general mood representing an entire feeling of the plurality of occupants, based on the estimated individual feeling of the each of the plurality of occupants; deciding a service to be provided to a group of the plurality of occupants, based on the estimated general mood; and outputting a command to provide the decided service.

Abstract: Systems and methods for adjusting user interface elements based on user accuracy and content consumption are disclosed. According to an aspect, a method includes receiving user input for interacting with one or more elements presented by a user interface. The method also includes determining an estimate of user accuracy for correctly selecting the one or more elements. Further, the method includes determining a rate of consumption of content presented by the user interface. The method also includes adjusting one of a position and size of at least one of the elements based on the estimate of user accuracy and the rate of consumption of content.

Abstract: An acoustic device that has a neck loop that is constructed and arranged to be worn around the neck. The neck loop includes a housing with a first acoustic waveguide having a first sound outlet opening, and a second acoustic waveguide having a second sound outlet opening. There is a first open-backed acoustic driver acoustically coupled to the first waveguide and a second open-backed acoustic driver acoustically coupled to the second waveguide.

Abstract: A method and apparatus for training a language model, include generating a first training feature vector sequence and a second training feature vector sequence from training data. The method is configured to perform forward estimation of a neural network based on the first training feature vector sequence, and perform backward estimation of the neural network based on the second training feature vector sequence. The method is further configured to train a language model based on a result of the forward estimation and a result of the backward estimation.