Abstract: A system and method for temporarily disabling keyword detection to avoid detection of machine-generated keywords. A local device may operate two keyword detectors. The first keyword detector operates on input audio data received by a microphone to capture keywords uttered by a user. In these instances, the keyword may be detected by the first detector and the audio data may be indicated for speech processing. The system may determine output audio data responsive to the input audio data. The local device may process the output audio data to determine that it also includes the keyword. The device may then disable the first keyword detector while the output audio data is played back by an audio speaker of the local device. Thus the local device may avoid detection of a keyword originating from the output audio. The first keyword detector may be reactivated after a time interval during which the keyword might be detectable in the output audio.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for speech recognition using neural networks. A feature vector that models audio characteristics of a portion of an utterance is received. Data indicative of latent variables of multivariate factor analysis is received. The feature vector and the data indicative of the latent variables is provided as input to a neural network. A candidate transcription for the utterance is determined based on at least an output of the neural network.

Abstract: A system and method for real-time fraud detection with a social engineering phoneme (SEP) watchlist of phoneme sequences may perform real-time fraud prevention operations including receiving incoming call interactions and grouping the call interactions into one or more clusters, each cluster associated with a speaker's voice based on voiceprints. For a pair of voiceprints in a cluster, a phoneme sequence is extracted for each voice print. From the extracted phoneme sequences, a similarity score is then calculated to determine if a match exists between the extracted phoneme sequences based on a threshold. If determined a match exists, the phoneme sequence may be added to a SEP watchlist.

Abstract: One embodiment provides a method, including: receiving, at an information handling device, text associated with a user command; storing, in a data store, an encrypted form of the text associated with the user command; determining, using a processor, whether the encrypted form of the text has been detected in other user commands in exceedance of a predetermined threshold; and storing, responsive to determining that the encrypted form of the text has been detected in the other user commands in exceedance of the predetermined threshold, an unencrypted transcript of the text in a data table. Other aspects are described and claimed.

Abstract: An incident report management system is configured to receive and analyze incident reports relating to workplace accidents and injuries. A natural language processing function utilizes word dictionaries of varying type and scope to reduce an incident description to a set of core components that are far smaller than the input text while also preserving important aspects of the input text. The reduced core component set may be analyzed for meaning, compared to large volumes of historic incident reports, and otherwise processed more quickly and more efficiently whether by an expert function or AI function. In this manner, the system is able to provide real-time feedback during submission of incidents to improve quality and completeness, and after submission of incidents to notify users of serious incidents, provide dashboard analytics, and identify underlying and undiscovered risks in the workplace.

Abstract: The technology relates to systems and methods for transcribing audio of a meeting. Upon transcribing the audio, the systems and methods can parse different portions of the prescribed audio so that they may attribute the different portions to a particular speaker. These transcribed portions that are attributed to a particular speaker are made available for viewing and interacting using a graphical user interface.

Abstract: A method for constructing a decoding network, a speech recognition method, a device, an apparatus, and a storage medium are provided. The method for constructing a decoding network includes: acquiring a general language model, a domain language model, and a general decoding network generated based on the general language model; generating a domain decoding network based on the domain language model and the general language model; and integrating the domain decoding network with the general decoding network to obtain a target decoding network. The speech recognition method includes: decoding to-be-recognized speech data by using a target decoding network to obtain a decoding path for the to-be-recognized speech data; and determining a speech recognition result for the to-be-recognized speech data based on the decoding path for the to-be-recognized speech data.

Abstract: Described herein is a method of encoding an audio signal. The method comprises: generating a plurality of subband audio signals based on the audio signal; determining a spectral envelope of the audio signal; for each subband audio signal, determining autocorrelation information for the subband audio signal based on an autocorrelation function of the subband audio signal; and generating an encoded representation of the audio signal, the encoded representation comprising a representation of the spectral envelope of the audio signal and a representation of the autocorrelation information for the plurality of subband audio signals. Further described are methods of decoding the audio signal from the encoded representation, as well as corresponding encoders, decoders, computer programs, and computer-readable recording media.

Abstract: Example methods and systems are directed to automatic data card generation for datasets. A data card is a summary that describes quantitative aspects of a dataset, qualitative aspects of a dataset, or both. The data samples and documentation of a dataset are analyzed automatically to determine a number of samples, a primary data type, a license, or any suitable combination thereof. Data formats for data and documentation of the dataset may be automatically recognized. Language of text data may be automatically recognized. The most frequent language for the text data may be identified as the primary language of the dataset. A data card may be created for the dataset. The data card may indicate the number of samples, the data formats used in the data set, the language of text data in the dataset, or any suitable combination thereof.

Abstract: Input context for a statistical dialog manager may be provided. Upon receiving a spoken query from a user, the query may be categorized according to at least one context clue. The spoken query may then be converted to text according to a statistical dialog manager associated with the category of the query and a response to the spoken query may be provided to the user.

Abstract: This disclosure relates generally to systems, methods, and computer readable media for providing improved insights and annotations to enhance recorded audio, video, and/or written transcriptions of testimony. For example, in some embodiments, a method is disclosed for correlating non-verbal cues recognized from an audio and/or video recording of testimony to the corresponding testimony transcript locations. In other embodiments, a method is disclosed for providing testimony-specific artificial intelligence-based insights and annotations to a testimony transcript, e.g., based on the use of machine learning, natural language processing, and/or other techniques. In still other embodiments, a method is disclosed for providing smart citations to a testimony transcript, e.g., which track the location of semantic constructs within the transcript over the course of various modifications being made to the transcript.

Abstract: Embodiments are provided for generating a reasonable language model learning for text data in a knowledge graph in a computing system by a processor. One or more data sources and one or more triples may be analyzed from a knowledge graph. Training data having one or more candidate labels associated with one or more of the triples may be generated. One or more reasonable language models may be trained based on the training data.

Abstract: An electronic apparatus, including a microphone; a memory configured to store at least one instruction; and a processor configured to: acquire a first token corresponding to a first user voice input in a first language acquired through the microphone, acquire a first text in a second language by inputting the first token into a first neural network model, acquire a feature value corresponding to a predicted subsequent token, which is predicted to be uttered after the first token, by inputting the first text into a second neural network model, and based on a second token being acquired subsequent to the first token, acquire a second text in the second language by inputting the first token, the second token, the first text, and the feature value into the first neural network model.

Abstract: In some embodiments, an exemplary inventive system for improving computer speed and accuracy of automatic speech transcription includes at least components of: a computer processor configured to perform: generating a recognition model specification for a plurality of distinct speech-to-text transcription engines; where each distinct speech-to-text transcription engine corresponds to a respective distinct speech recognition model; receiving at least one audio recording representing a speech of a person; segmenting the audio recording into a plurality of audio segments; determining a respective distinct speech-to-text transcription engine to transcribe a respective audio segment; receiving, from the respective transcription engine, a hypothesis for the respective audio segment; accepting the hypothesis to remove a need to submit the respective audio segment to another distinct speech-to-text transcription engine, resulting in the improved computer speed and the accuracy of automatic speech transcription and gen

Abstract: A method for implementing a privacy-preserving automatic speech recognition system using federated learning. The method includes receiving, from respective client devices, at a cloud server, local acoustic model weights for a neural network-based acoustic model of a local automatic speech recognition system running on the respective client devices, wherein the local acoustic model weights are generated at the respective client devices without labelled data, updating a global automatic speech recognition system based on (a) the local acoustic model weights received from the respective client devices and (b) global acoustic model weights of the global automatic speech recognition system derived from labelled data to obtain an updated global automatic speech recognition system, and sending the updated global automatic speech recognition system to the respective client devices to operate as a new local automatic speech recognition system.

Abstract: A method of presenting prompt information by utilizing a neural network which includes a BERT model and a graph convolutional neural network (GCN), comprising: generating a first vector based on a combination of an entity, a context of the entity, a type of the entity and a part of speech of the context by using BERT model; generating a second vector based on each of predefined concepts by using BERT model; generating a third vector based on a graph which is generated based on the concepts and relationships thereamong, by using GCN; generating a fourth vector by concatenating the second and third vectors; calculating semantic similarity between the entity and each concept based on the first and fourth vectors; determining, based on the first vector and the semantic similarity, that the entity corresponds to one of the concepts; and generating the prompt information based on the determined concept.

Abstract: One or more computer processors obtain an initial subnetwork at a target sparsity and an initial pruning mask from a pre-trained self-supervised learning (SSL) speech model. The one or more computer processors finetune the initial subnetwork, comprising: the one or more computer processors zero out one or more masked weights in the initial subnetwork specified by the initial pruning mask; the one or more computer processors train a new subnetwork from the zeroed out subnetwork; the one or more computer processors prune one or more weights of lowest magnitude in the new subnetwork regardless of network structure to satisfy the target sparsity. The one or more computer processors classify an audio segment with the finetuned subnetwork.

Abstract: Provided is a computer-implemented method, system, and computer program product for process mining asynchronous support conversations using attributed directly follows graphing. A processor may collect a plurality of conversation threads from an asynchronous data stream. The processor may label each utterance of a plurality of utterances from the plurality of conversation threads with an event label. The processor may analyze the event label for each utterance of the plurality of utterances. The processor may generate, based on the analyzing of the event label for each utterance, an attributed directly follows graph (DFG).

Abstract: An example system includes a processor to receive a summary of a conversation to be generated. The processor can input the summary into a trained summary-grounded conversation generator. The processor can receive a generated conversation from the trained summary-grounded conversation generator.

Abstract: An audio signal encoding method and device are provided. The method and device are used to encode an audio signal to obtain a bitstream representing the analog audio signal, in which a proper bit allocation for spectral coefficients can be performed.