Abstract: Techniques for updating a machine learning (ML) model are described. A device or system may receive input data corresponding to a natural or non-natural language (e.g., gesture) input. Using a first ML model, the device or system may determine the input data corresponds to a data category of a plurality of data categories. Based on the data category, the device or system may select a ML training type from among a plurality of ML training types. Using the input data, the device or system may perform the selected ML training type with respect to a runtime ML model to generate an updated ML model.

Abstract: A method of generating a residual signal performed by an encoder includes identifying an input signal including an audio sample, generating a first residual signal from the input signal using linear predictive coding (LPC), generating a second residual signal having a less information amount than the first residual signal by transforming the first residual signal, transforming the second residual signal into a frequency domain, and generating a third residual signal having a less information amount than the second residual signal from the transformed second residual signal using frequency-domain prediction (FDP) coding.

Abstract: Mitigating mistranscriptions resolves errors in a transcription of the audio portion of a video based on a semantic matching with contextualized data electronically garnered from one or more sources other than the audio portion of the video. A mistranscription is identified using a pretrained word embedding model that maps words to an embedding space derived from the contextualizing data. A similarity value for each vocabulary word of a multi-word vocabulary of the pretrained word embedding model is determined in relation to the mistranscription. Candidate words are selected based on the similarity values, each indicating a closeness of a corresponding vocabulary word to the mistranscription. The textual rendering is modified by replacing the mistranscription with a candidate word that, based on average semantic similarity values, is more similar to the mistranscription than is each other candidate word.

Abstract: Methods and systems are described for generating dynamic conversational responses using machine learning models. The dynamic conversational responses may be generated in real time and reflect the likely goals and/or intents of a user. The machine learning model may provide these features by monitoring one or more user actions and/or lengths of time between one or more user actions during conversational interactions.

Abstract: A document is divided into sections based on a characteristic of the text in the document. Characteristics may include specific characters such as paragraph breaks or selected punctuation, the topics or categories of the text, or a quantity of text in each section. Each section of the document may be represented by an embedding vector in a semantic embedding space. Values are determined based on the text in each section and the semantic characteristics of each section, such as word order, capitalization, punctuation, and word meaning. When a query is received, a vector value representing the query is determined based on the text and semantic characteristics of the query. Based on the similarity between the values determined for the query and those determined for the sections of a document, the specific section of a potentially large document that most closely matches the query is determined and included in a response.

Abstract: Various embodiments of an apparatus, methods, systems and computer program products described herein are directed to a Voice Analysis Engine. According to various embodiments, the Voice Analysis Engine receives first streaming prompt data from a computing device. The Voice Analysis Engine analyzes the first streaming prompt data to provide feedback to the user of the computing device. Upon determining the first streaming prompt data satisfies one or more criteria, the Voice Analysis Engine receives second streaming prompt data from the computing device. The Voice Analysis Engine analyzes the streaming prompt data to predict a respiratory state of the user of the computing device.

Abstract: An audio processing method includes the following operations. A calculated value is obtained according to multiple audio clock frequency information contained in multiple audio input packets. An audio sampling frequency is generated according to the calculated value and a link symbol clock signal. Multiple audio output packets corresponding to the audio input packets are generated according to the audio sampling frequency.

Abstract: Methods and systems for speech recognition in an aircraft are disclosed. Methods and systems include executing an air traffic control transcription application using first acoustic and language models and executing a command and control speech recognition application using second acoustic and language models. Flight context data is processed to identify additional words not included in training of the second acoustic and language model but included in training of the first acoustic and language models. Acoustic and language model parameters are extracted corresponding to the additional words from the first acoustic and language models. The extracted acoustic and language model parameters are added to the second acoustic and language models. An aircraft control command is generated that encapsulates at least one of the additional words using the command and control speech recognition application.

Abstract: A method includes receiving an input audio signal corresponding to utterances spoken by multiple speakers. The method also includes encoding the input audio signal into a sequence of T temporal embeddings. During each of a plurality of iterations each corresponding to a respective speaker of the multiple speakers, the method includes selecting a respective speaker embedding for the respective speaker by determining a probability that the corresponding temporal embedding includes a presence of voice activity by a single new speaker for which a speaker embedding was not previously selected during a previous iteration and selecting the respective speaker embedding for the respective speaker as the temporal embedding. The method also includes, at each time step, predicting a respective voice activity indicator for each respective speaker of the multiple speakers based on the respective speaker embeddings selected during the plurality of iterations and the temporal embedding.

Abstract: This application discloses an information conversion method and apparatus, a storage medium, and an electronic apparatus.

Abstract: Systems and methods of predicting glottal insufficiency by at least one hardware processor including receiving a voice recording comprising a phonation by a subject, analysis of the voice recording to calculate a fundamental frequency contour curve of the phonation, and measurement of at least one of (i) a time period from a start of the phonation until the contour curve reaches a settled level, (ii) a slope of the contour curve during the time period, and (iii) an area under the contour curve during that time period. In certain embodiments, the processor subsequently, determines a glottal closure insufficiency in the subject based on these measurements.

Abstract: Systems and methods for automatic evaluation of the quality of NLG outputs. In some aspects of the technology, a learned evaluation model may be pretrained first using NLG model pretraining tasks, and then with further pretraining tasks using automatically generated synthetic sentence pairs. In some cases, following pretraining, the evaluation model may be further fine-tuned using a set of human-graded sentence pairs, so that it learns to approximate the grades allocated by the human evaluators.

Abstract: Described are systems, methods, and apparatus that detect keywords in one or more speech segments to authenticate that the speech is generated by the speaker as part of an intentional enrollment by the speaker into a service. For example, as a speech segment is received as part of an enrollment process, the speech segment may be converted into a log melspectrogram and the log melspectrogram may be processed using a machine learning model to determine if an expected keyword is represented by the log melspectrogram. If the keyword is detected, it may be determined that the speech output by the speaker is output as part of an intentional enrollment process.

Abstract: Techniques are provided to validate a digitized audio signal that is generated by a conference participant. Reference speech features of the conference participant are obtained, either via samples provided explicitly by the participant, or collected passively via prior conferences. The speech features include one or more of word choices, filler words, common grammatical errors, idioms, common phrases, pace of speech, or other features. The reference speech features are compared to features observed in the digitized audio signal. If the reference speech features are sufficiently similar to the observed speech features, the digitized audio signal is validated and the conference participant is allowed to remain in the conference. If the validation is not successful, a variety of possible actions are taken, including alerting an administrator and/or terminating the participant's attendance in the conference.

Abstract: A computer implemented method, system and program product is provided for linguistic alignment in specific user targeted messaging. In one embodiment, new and previously existing data about a specific user is analyzed and personality insights are determined. Location of the user is also determined. Using this location and collected data and personality insights, a multilayered set of linguistic preferences is determined for the specific user. This set is used to customize a message for the specific user based on the linguistic set and ultimately a message is sent to the specific user using a selected messaging channel.

Abstract: Systems, computer-implemented methods, and computer program products to facilitate a dual learning bridge between text and a knowledge graph are provided. According to an embodiment, a system can comprise a processor that executes computer executable components stored in memory. The computer executable components comprise a model component that employs a model to learn associations between text data and a knowledge graph. The computer executable components further comprise a translation component that uses the model to bidirectionally translate second text data and one or more knowledge graph paths based on the associations.

Abstract: Devices, systems, and methods are provided for context-based abusive language detection and responses. A method may include identifying text associated with first video content, and determining that a first word in the text matches a first keyword indicative of abusive language. The method may include determining a first label associated with the first word, the first label indicating that the first word is ambiguous. The method may include identifying a first sentence of the text, the first sentence including the first word. The method may include determining first and second context of the first word and the first sentence. The method may include determining, based on the first and second context, using a machine learning model, a second label associated with the first sentence, the second label indicating a probability that the first sentence includes abusive language. The method may include generating second video content for presentation.

Abstract: A system and method for processing an audio input signal includes a microphone, a controller, and a communication link that may be coupled to a remote speaker. The microphone captures the audio input signal and communicates the audio input signal to the controller, and the controller is coupled to the communication link. The controller includes executable code to generate, via a linear noise reduction filtering algorithm, a first resultant based upon the audio input signal, and generate, via non-linear post filtering algorithm, a second resultant based upon the first resultant. An audio output signal is generated based upon the second resultant employing a feature restoration algorithm. The audio output signal is communicated, via the communication link, to a speaker that may be at a remote location.

Abstract: A method for training a speech-emotion recognition classifier under a continuously updatable and re-trainable ASER machine learning model. The quantified training data is generated by first processing the utterances of a human speech source and the associated texts in an emotion evaluation and rating process with normalization; then, extracting the features of the utterance; quantifying the feature attributes of the extracted features by labelling, tagging, and weighting the feature attributes, with their values assigned under measurable scales. The quantified training data comprises the normalized results of the emotion evaluation and rating process, the extracted features, the quantified emotional feature attributes, and the hash values of the quantified emotional feature attributes.

Abstract: A method, computer system, and a computer program product for natural language processing are provided. A first text corpus that includes semi-structured content that includes hierarchical nodes may be received. Some of the hierarchical nodes may be masked. Node embeddings and level embeddings may be generated from the semi-structured content of the first text corpus and from the masked hierarchical nodes. The node embeddings and the level embeddings may be included in a bi-directional transformer model. The bi-directional transformer model may be trained on the first text corpus by reducing loss from the bi-directional transformer model predicting the masked hierarchical nodes.