Abstract: The embodiment of the disclosure provides a method for voice controlling, a terminal device, a cloud server and a system. The method includes: receiving voice information that the user performs voice controlling on a terminal device; transmitting voice information to the cloud server, so that the cloud server determines, according to the voice information, a voice control and a control instruction that match the voice information in the current interface, and generates a corresponding voice control instruction; receiving the voice control instruction transmitted by the cloud server; and controlling, according to the voice control instruction, a corresponding voice control of the terminal device to perform an operation. The method of the embodiments of the present disclosure achieves controlling over the controls in the interface through the voice, which deepens the controlling degree of the voice over the terminal device, and improves the user experience.

Abstract: Disclosed are systems, methods, and non-transitory computer-readable media for probabilistic word embeddings for text classification. A text classification system receives a message including a keyword and determines an embedding probability distribution representing the keyword. The text classification system then determines an embedding value for the keyword based on the embedding probability distribution. The text classification system uses the embedding value as input into a set of mathematical functions, yielding a first set of coefficient values for the keyword. Each respective mathematical function from the set corresponds to a respective classification label from a set of classification labels and defines a continuous surface.

Abstract: A control module, for lighting or other environmental control or the like has embedded voice responsive control capability. The recognizable command set supported by the control module is customizable. Software for a data processor configures the control module to identify a user in the vicinity of the control module and change voice response software based at least in part on identity of the user. The changed voice response software configures the control module, to recognize a customized set of vocal commands, e.g. associated with the identified user, in response to speech detected via a microphone and an audio processor of the control module, without requiring network communication. The control module issues control signals to equipment controlled by the module that correspond to the recognized commands of the customized vocal command set.

Abstract: A technology is described for providing virtual call assistants and call options to identified callers. An example method may include identifying a caller and a call recipient using addressing information included in a call received at a call control service. After identifying the caller, a priority designation assigned to the caller may be obtained from an agent linking profile. An interactive agent linked to the priority designation assigned to the caller may be identified, wherein the interactive agent may be configured to provide prioritized call services. Thereafter, the interactive agent linked to the priority designation may be invoked.

Abstract: An exemplary automatic speech recognition (ASR) system may receive an audio input including a segment of speech. The segment of speech may be independently processed by general ASR and domain-specific ASR to generate multiple ASR results. A selection between the multiple ASR results may be performed based on respective confidence levels for the general ASR and domain-specific ASR. As incremental ASR is performed, a composite result may be generated based on general ASR and domain-specific ASR.

Abstract: Embodiments are associated with a speaker-independent acoustic model capable of classifying senones based on input speech frames and on first parameters of the speaker-independent acoustic model, a speaker-dependent acoustic model capable of classifying senones based on input speech frames and on second parameters of the speaker-dependent acoustic model, and a discriminator capable of receiving data from the speaker-dependent acoustic model and data from the speaker-independent acoustic model and outputting a prediction of whether received data was generated by the speaker-dependent acoustic model based on third parameters.

Abstract: An end detector configured to receive the feature data and detect an end point of a keyword, and a start detector configured to receive an indication of the detected end point and process the feature data associated with corresponding input frames to detect a start point of the keyword. The start detector and end detector comprise neural networks trained through a process using a cross-entropy cost function for non-Region of Target (ROT) frames and a One-Spike Connectionist Temporal Classification cost function for ROT frames.

Abstract: A classification training system for binary and multi-class classification comprises a neural network operable to perform classification of input data, a training dataset including pre-segmented, labeled training samples, and a classification training module operable to train the neural network using the training dataset. The classification training module includes a forward pass processing module, and a backward pass processing module. The backward pass processing module is operable to determine whether a current frame is in a region of target (ROT), determine ROT information such as beginning and length of the ROT and update weights and biases using a cross-entropy cost function and One Spike Connectionist Temporal Classification (OSCTC) cost function. The backward pass module further computes a soft target value using ROT information and computes a signal output error using the soft target value and network output value.

Abstract: The systems and methods of the present disclosure generally relate to a data processing system that can identify and surface alternative requests when presented with ambiguous, unclear, or other requests to which a data processing system may not be able to respond. The data processing system can improve the efficiency of network transmissions to reduce network bandwidth usage and processor utilization by selecting alternative requests that are responsive to the intent of the original request.

Abstract: An acoustic event detection and classification system includes a start-end point detector and multi-class acoustic event classification. A classification training system comprises a neural network configured to perform classification of input data, a training dataset including pre-segmented, labeled training samples, and a classification training module configured to train the neural network using the training dataset. The classification training module includes a forward pass processing module, and a backward pass processing module. The backward pass processing module is configured to determine whether a current frame is in a region of target (ROT), determine ROT information such as beginning and length of the ROT and update weights and biases using a cross-entropy cost function and a many-or-one detection (MOOD) cost function. The backward pass module further computes a soft target value using ROT information and computes a signal output error using the soft target value and network output value.

Abstract: Described herein is a system for media presence detection in audio. The system analyzes audio data to recognize whether a given audio segment contains sounds from a media source as a way of differentiating recorded media source sounds from other live sounds. In exemplary embodiments, the system includes a hierarchical model architecture for processing audio data segments, where individual audio data segments are processed by a trained machine learning model operating locally, and another trained machine learning model provides historical and contextual information to determine a score indicating the likelihood that the audio data segment contains sounds from a media source.

Abstract: A system includes at least one processor to perform natural language processing on text from at least one document and assign the at least one document to at least one subjective wellbeing dimension by comparing the text from the at least one document with a subjective wellbeing dimension filter for each subjective wellbeing dimension, insert the at least one document into at least one bin, each bin associated with a particular subjective wellbeing dimension, and analyze each document in each bin associated with the particular subjective wellbeing dimension to determine a score for each subjective wellbeing dimension and an overall score that is based on each score for each subjective wellbeing dimension.

Abstract: A device includes a processor configured to, in response to determining that an input phrase includes a first term that is included in a term hierarchy, generate a second phrase by replacing the first term in the input phrase with a second term included in the term hierarchy. The processor is configured to determine that interactive response (IR) training data indicates that the input phrase is associated with a user intent indicator. The processor is configured to determine that user interaction data indicates that a first proportion of user phrases received by an IR system correspond to the user intent indicator. The processor is configured to update speech-to-text training data based on the input phrase and the second phrase so that a second proportion of training phrases of the speech-to-text training data correspond to the user intent indicator. The second proportion is based on the first proportion. A speech-to-text model is trained based on the speech-to-text training data.

Abstract: Embodiments of the present invention provide systems, methods, and computer storage media directed to providing talk back automation for applications installed on a mobile device. To do so actions (e.g., talk back features) can be created, via the digital assistant, by recording a series of events that are typically provided by a user of the mobile device when manually invoking the desired action. At a desired state, the user may select an object that represents the output of the application. The recording embodies the action and can be associated with a series of verbal commands that the user would typically announce to the digital assistant when an invocation of the action is desired. In response, the object is verbally communicated to the user via the digital assistant, a different digital assistant, or even another device. Alternatively, the object may be communicated to the same application or another application as input.

Abstract: A method for training hotword detection includes receiving a training input audio sequence including a sequence of input frames that define a hotword that initiates a wake-up process on a device. The method also includes feeding the training input audio sequence into an encoder and a decoder of a memorized neural network. Each of the encoder and the decoder of the memorized neural network include sequentially-stacked single value decomposition filter (SVDF) layers. The method further includes generating a logit at each of the encoder and the decoder based on the training input audio sequence. For each of the encoder and the decoder, the method includes smoothing each respective logit generated from the training input audio sequence, determining a max pooling loss from a probability distribution based on each respective logit, and optimizing the encoder and the decoder based on all max pooling losses associated with the training input audio sequence.

Abstract: Of the four primary approaches to processing language by computer, only the parsing approach considers the semantic and syntactic components from the start. In doing so, however, the required resources expand rapidly as the scope of the language processed increases. And as that scope increases, the performance of parsing systems decreases. A natural language processor uses a tumbling-frequency phrase chain parser as described herein which circumvents this resource-intensive step in parsing, while quickly and almost effortlessly arriving at the next step in natural-language processing with far more accurate results involving a partitioning dictionary and phrase chains.

Abstract: A method includes receiving information from each device in multiple devices of one user, where the information includes a device type, determining, according to at least the received information, a device responding to a voice instruction in the multiple devices, and sending a message to a device different from the device responding to a voice instruction in the multiple devices, where the message is used to instruct a voice assistant of the different device not to respond to the voice instruction.

Abstract: Described techniques select portions of an audio stream for transmission to a trained machine learning application, which generates response recommendations in real-time. This real-time response is facilitated by the system identifying, selecting and transmitting those portions of the audio stream likely to be most relevant to the conversation. Portions of an audio stream less likely to be relevant to the conversation are identified accordingly and not transmitted. The system may identify the relevant portions of an audio stream by detecting events in a contemporaneous event stream, use a trained machine learning model to identify events in an audio stream, or both.

Abstract: Systems and methods for intelligently training a subject machine learning model includes identifying new observations comprising a plurality of distinct samples unseen by a target model during a prior training; creating an incremental training corpus based on randomly sampling a collection of training data samples that includes a plurality of new observations and a plurality of historical training data samples used in the prior training of the target model; implementing a first training mode that includes an incremental training of the target model using samples from the incremental training corpus as model training input; computing performance metrics of the target model based on the incremental training; evaluating the performance metrics of the target model against training mode thresholds; and selectively choosing based on the evaluation one of maintaining the first training mode and automatically switching to a second training mode that includes a full retraining of the target model.

Abstract: A method of generating text using an adversarial network includes receiving a limited dataset. The limited dataset includes real data having actual parameters and actual sentences. The method includes receiving content data that includes a concept related to a portion of the real data or that causes an issue of the real data. The method includes generating relationships between the real data and the content data. The method includes embedding the content data with the real data in an encoder output that includes content vector embedding. The method includes generating an additional parameter set that includes additional parameters and one or more additional statements. The additional parameter set may be supplemental to the real data and configured to enhance an expressiveness of a model. The method includes generating explanatory statement based on the additional parameter set and the relationships.