Abstract: Methods and systems are disclosed for improving dialog management for task-oriented dialog systems. The disclosed dialog builder leverages machine teaching processing to improve development of dialog managers. In this way, the dialog builder combines the strengths of both rule-based and machine-learned approaches to allow dialog authors to: (1) import a dialog graph developed using popular dialog composers, (2) convert the dialog graph to text-based training dialogs, (3) continuously improve the trained dialogs based on log dialogs, and (4) generate a corrected dialog for retraining the machine learning.

Abstract: A speech-processing system may provide access to one or more virtual assistants via a voice-controlled device. A user may leverage a first virtual assistant to translate a natural language command from a first language into a second language, which the device can forward to a second virtual assistant for processing. The device may receive a command from a user and send input data representing the command to a first speech-processing system representing the first virtual assistant. The device may receive a response in the form of a first natural language output from the first speech-processing system along with an indication that the first natural language output should be directed to a second speech-processing system representing the second virtual assistant. For example, the command may be in the first language, and the first natural language output may be in the second language, which is understandable by the second speech-processing system.

Abstract: Systems and methods are described for generating a dynamic label for a real-time communication session. An ongoing communication session is monitored to identify a content characteristic of the communication session. A size of a sliding window is determined based on the content characteristic, where the size of the sliding window defines a segment of the communication session to include in the most recent subset of communications. The most recent subset of communications is analyzed to identify relevant words based on one or more relevancy criteria. A dynamic label associated with the communication session is generated, where the dynamic label includes at least a selected one of the relevant words.

Abstract: Various embodiments set forth systems and techniques for explaining domain-specific terms detected in a media content stream. The techniques include detecting a speech portion included in an audio signal; determining that the speech portion comprises a domain-specific term; determining an explanatory phrase associated with the domain-specific term; and integrating the explanatory phrase associated with the domain-specific term into playback of the audio signal.

Abstract: Embodiments described herein provides a training mechanism that transfers the knowledge from a trained BERT model into a much smaller model to approximate the behavior of BERT. Specifically, the BERT model may be treated as a teacher model, and a much smaller student model may be trained using the same inputs to the teacher model and the output from the teacher model. In this way, the student model can be trained within a much shorter time than the BERT teacher model, but with comparable performance with BERT.

Abstract: An approach is described with respect to dynamically constructing and configuring a conversational agent learning model. Various aspects of the conversational agent learning model may be constructed and updated without continuous intervention of a domain administrator. A method pertaining to such approach may include retrieving a corpus of information. The corpus of information may include records from a set of repositories and external data, including data from social networks or applications. The method further may include configuring the conversational agent learning model based upon the retrieved corpus of information. The method further may include deploying the conversational agent learning model by facilitating interaction between the conversational agent learning model and a plurality of clients. The method further may include updating the conversational agent learning model to address any modification to the corpus of information.

Abstract: Methods and systems for converting an input content item into an output content item to enhance comprehension of the message by an interlocutor, based on contexts. For example, the conversion may occur in any message service: when an interlocutor writes a message in English (or any other language), he or she might include a regional dialect (purposively or not), such as a piece of slang, that the other interlocutors may not understand, although they all generally write and understand English. In such circumstances, the regional dialect is identified and replaced with either a more globalized word or with another linguistic regionalism that is understandable for the intended interlocutor.

Abstract: Techniques are described and relate to providing computing services using embeddings of a transformer-based encoder. In an example, a computer system generates, by using a machine learning (ML) transformer, an embedding vector based at least in part on text. The computer system stores the embedding vector and an association between the embedding vector and the text in a data store. Further, the computer system determines that a task is to be performed based at least in part on natural language understanding (NLU) of the text. The computer system receives the embedding vector from the data store based at least in part on the association between the embedding vector and the text. The task is performed based at least in part on the embedding vector after being received from the data store.

Abstract: Techniques are described herein for identifying a failed hotword attempt. A method includes: receiving first audio data; processing the first audio data to generate a first predicted output; determining that the first predicted output satisfies a secondary threshold but does not satisfy a primary threshold; receiving second audio data; processing the second audio data to generate a second predicted output; determining that the second predicted output satisfies the secondary threshold but does not satisfy the primary threshold; in response to the first predicted output and the second predicted output satisfying the secondary threshold but not satisfying the primary threshold, and in response to the first spoken utterance and the second spoken utterance satisfying one or more temporal criteria relative to one another, identifying a failed hotword attempt; and in response to identifying the failed hotword attempt, providing a hint that is responsive to the failed hotword attempt.

Abstract: Implementations can receive audio data corresponding to a spoken utterance of a user, process the audio data to generate a plurality of speech hypotheses, determine an action to be performed by an automated assistant based on the speech hypotheses, and cause the computing device to render an indication of the action. In response to the computing device rendering the indication, implementations can receive additional audio data corresponding to an additional spoken utterance of the user, process the additional audio data to determine that a portion of the spoken utterance is similar to an additional portion of the additional spoken utterance, supplant the action with an alternate action, and cause the automated assistant to initiate performance of the alternate action. Some implementations can determine whether to render the indication of the action based on a confidence level associated with the action.

Abstract: Systems and methods are disclosed for displaying electronic multimedia content to a user. One computer-implemented method for manipulating electronic multimedia content includes generating, using a processor, a speech model and at least one speaker model of an individual speaker. The method further includes receiving electronic media content over a network; extracting an audio track from the electronic media content; and detecting speech segments within the electronic media content based on the speech model. The method further includes detecting a speaker segment within the electronic media content and calculating a probability of the detected speaker segment involving the individual speaker based on the at least one speaker model.

Abstract: Embodiments of the disclosed technologies incorporate taxonomy information into a cross-lingual entity graph and input the taxonomy-informed cross-lingual entity graph into a graph neural network. The graph neural network computes semantic alignment scores for node pairs. The semantic alignment scores are used to determine whether a node pair represents a valid machine translation.

Abstract: A virtual assistant server determines at least one user intent based on an analysis of a received conversational user input. One or more of a plurality of views is identified based on the at least one user intent. Further, the virtual assistant server retrieves content based on the at least one user intent or the identified one or more views. The virtual assistant server determines one of a plurality of graphical user interface layers to display for each of one or more parts of the content and the identified one or more views based at least on one or more factors related to the content. Subsequently, the virtual assistant server outputs instructions based on the determined one of the graphical user interface layers in response to the received conversational user input.

Abstract: Techniques are described herein for multi-factor audio watermarking. A method includes: receiving audio data; processing the audio data to generate predicted output that indicates a probability of one or more hotwords being present in the audio data; determining that the predicted output satisfies a threshold that is indicative of the one or more hotwords being present in the audio data; in response to determining that the predicted output satisfies the threshold, processing the audio data using automatic speech recognition to generate a speech transcription feature; detecting a watermark that is embedded in the audio data; and in response to detecting the watermark: determining that the speech transcription feature corresponds to one of a plurality of stored speech transcription features; and in response to determining that the speech transcription feature corresponds to one of the plurality of stored speech transcription features, suppressing processing of a query included in the audio data.

Abstract: The present disclosure discloses a method and apparatus for training a model, a method and apparatus for synthesizing a speech, a device and a storage medium, and relates to the field of natural language processing and deep learning technology. The method for training a model may include: determining a phoneme feature and a prosodic word boundary feature of sample text data; inserting a pause character into the phoneme feature according to the prosodic word boundary feature to obtain a combined feature of the sample text data; and training an initial speech synthesis model according to the combined feature of the sample text data, to obtain a target speech synthesis model.

Abstract: The present disclosure provides an information switching method. The method includes: obtaining tilting information after an tilt direction of a device changes; searching a pre-set tilt direction matching the tilting information and determining pre-set information corresponding to the matched pre-set tilt direction; and switching first input information of the device to second input information, where the second input information is determined based on the pre-set information matching the pre-set tilt direction.

Abstract: A method of training a speech model includes receiving, at a voice-enabled device, a fixed set of training utterances where each training utterance in the fixed set of training utterances includes a transcription paired with a speech representation of the corresponding training utterance. The method also includes sampling noisy audio data from an environment of the voice-enabled device. For each training utterance in the fixed set of training utterances, the method further includes augmenting, using the noisy audio data sampled from the environment of the voice-enabled device, the speech representation of the corresponding training utterance to generate noisy audio samples and pairing each of the noisy audio samples with the corresponding transcription of the corresponding training utterance. The method additionally includes training a speech model on the noisy audio samples generated for each speech representation in the fixed set of training utterances.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media for evaluating and improving live translation captioning systems. An exemplary method includes: displaying a word in a first language; receiving a first audio sequence, the first audio sequence comprising a verbal description of the word; generating a first translated text in a second language; displaying the first translated text; receiving a second audio sequence, the second audio sequence comprising a guessed word based on the first translated text; generating a second translated text in the first language; determining a matching score between the word and the second translated text; determining a performance score of the live translation captioning system based on the matching score.

Abstract: Embodiments of this application include an information conversion method for translating source information. The source information is encoded to obtain a first code. A preset conversion condition is obtained. The preset conversion condition indicates a mapping relationship between the source information and a conversion result. The first code is decoded according to the source information, the preset conversion condition, and translated information to obtain target information. The target information and the source information are in different languages. Further, the translated information includes a word obtained through conversion of the source information into a language of the target information.

Abstract: A method for generating a comfort noise (CN) parameter is provided. The method includes receiving an audio input; detecting, with a Voice Activity Detector (VAD), a current inactive segment in the audio input; as a result of detecting, with the VAD, the current inactive segment in the audio input, calculating a CN parameter CNused; and providing the CN parameter CNused to a decoder. The CN parameter CNused is calculated based at least in part on the current inactive segment and a previous inactive segment.