Abstract: A method executed by a computing device includes determining a set of identigens for each phrase word of a phrase to produce sets of identigens. A set of identigens of the sets of identigens represents one or more different meanings of a phrase word of the phrase. The method further includes obtaining inflection information for one or more phrase words of the phrase. The method further includes selecting an identigen of a first set of identigens based on the inflection information to produce a first identigen selection for the first set of identigens having a selected meaning of one or more different meanings of the first phrase word. The method further includes interpreting remaining sets of identigens of the sets of identigens to produce an entigen group so that the entigen group represents a most likely meaning interpretation of the phrase.

Abstract: A method and system for detecting and localizing a target audio event in an audio clip is disclosed. The method and system use utilizes a hierarchical approach in which a dilated convolutional neural network to detect the presence of the target audio event anywhere in an audio clip based on high level audio features. If the target audio event is detected somewhere in the audio clip, the method and system further utilizes a robust audio vector representation that encodes the inherent state of the audio as well as a learned relationship between state of the audio and the particular target audio event that was detected in the audio clip. A bi-directional long short term memory classifier is used to model long term dependencies and determine the boundaries in time of the target audio event within the audio clip based on the audio vector representations.

Abstract: The present disclosure provides a method for denoising voice data, an electronic device, and a computer readable storage medium. The present disclosure relates to the technical field of artificial intelligence, such as Internet of Vehicles, smart cockpit, smart voice, and voice recognition. A specific embodiment of the method includes: receiving an input to-be-played first piece of voice data; and invoking, in response to not detecting a synthetic voice interruption signal in a process of playing the first piece of voice data, a preset first denoising algorithm to filter out noise data except for the first piece of voice data.

Abstract: A method for transcription is performed by a computer. The method includes: accepting input of a voice after causing a display unit to display a sentence including a plurality of words; acquiring first sound information being information concerning sounds corresponding to the sentence; acquiring second sound information being information concerning sounds of the voice accepted in the accepting; specifying a portion in the first sound information having a prescribed similarity to the second sound information; and correcting a character string in the sentence corresponding to the specified portion based on a character string corresponding to the second sound information.

Abstract: An encoder for encoding an audio signal has: an analyzer configured for deriving prediction coefficients and a residual signal from an unvoiced frame of the audio signal; a gain parameter calculator configured for calculating a first gain parameter information for defining a first excitation signal related to a deterministic codebook and for calculating a second gain parameter information for defining a second excitation signal related to a noise-like signal for the unvoiced frame; and a bitstream former configured for forming an output signal based on an information related to a voiced signal frame, the first gain parameter information and the second gain parameter information.

Abstract: A training device changes feedback formant frequencies which are formant frequencies of a picked-up speech signal, applies a lowpass filter, converts the picked-up speech signal, adds high-pass noise to the converted speech signal, feeds back the converted speech signal with the high-pass noise added to a subject, calculates a compensatory response vector by using pickup formant frequencies which are formant frequencies of a speech signal acquired by picking up an utterance made by the subject while feeding back a speech signal that has been converted with change of the feedback formant frequencies to the subject, and pickup formant frequencies which are formant frequencies of a speech signal acquired by picking up an utterance made by the subject while feeding back a speech signal that has been converted without change of the feedback formant frequencies to the subject, and determines an evaluation based on the compensatory response vector and a correct compensatory response vector.

Abstract: An apparatus for speech model with personalization via ambient context harvesting, is described herein. The apparatus includes a microphone, context harvesting module, confidence module, and training module. The context harvesting module is to determine a context associated with the captured audio signals. A confidence module is to determine a confidence of the context as applied to the audio signals. A training module is to train a neural network in response to the confidence being above a predetermined threshold.

Abstract: This invention provides a signal processing apparatus capable of obtaining an output signal of sufficiently high quality if the phase of an input signal is largely different from the phase of a true voice. The signal processing apparatus includes a voice detector that receives a mixed signal including a voice and a signal other than the voice and obtains existence of the voice as a voice flag, a corrector that receives the mixed signal and the voice flag and obtains a corrected mixed signal generated by correcting the mixed signal in accordance with a state of the voice flag, and a shaper that receives the corrected mixed signal and shapes the corrected mixed signal.

Abstract: Embodiments of this application disclose a bandwidth extension (BWE) method and apparatus. The method is performed by an electronic device, and includes: performing a time-frequency transform on a to-be-processed narrowband signal to obtain a corresponding initial low-frequency spectrum; obtaining a correlation parameter of a high-frequency portion and a low-frequency portion of a target broadband spectrum based on the initial low-frequency spectrum by using a neural network model; obtaining an initial high-frequency spectrum based on the correlation parameter and the initial low-frequency spectrum; and obtaining a broadband signal according to a target low-frequency spectrum and a target high-frequency spectrum.

Abstract: A control device of voice distribution including: at least one voice processing module arranged to—receive as input an audio signal including a first vocal message, and—provide as output an audio signal including a second vocal message, the first and second vocal messages being different one from the other and the second vocal message resulting from a processing of the first vocal message; a communication module arranged to establish and simultaneously manage a wireless, bidirectional and audio link with each one of a plurality of auxiliary devices, each link being connected to the input and/or the output of at least one voice processing module.

Abstract: Example embodiments relate to techniques for training artificial neural networks or oilier machine-learning encoders to accurately predict the pitch of input audio samples in a semitone or otherwise logarithmically-scaled pitch space. An example method may include generating, from a sample of audio data, two training samples by applying two different pitch shifts to the sample of audio training data. This can be done by converting the sample of audio data into the frequency domain and then shifting the transformed data. These known shifts are then compared to the predicted pitches generated by applying the two training samples to the encoder. The encoder is then updated based on the comparison, such that the relative pitch output by the encoder is improved with respect to accuracy. One or more audio samples, labeled with absolute pitch values, can then be used to calibrate the relative pitch values generated by the trained encoder.

Abstract: A keyword detection method includes: obtaining an enhanced speech signal of a to-be-detected speech signal, the enhanced speech signal corresponding to a target speech speed; performing speed adjustment on the enhanced speech signal to obtain a first speed-adjusted speech signal having a first speech speed, the first speech speed being different from the target speech speed; obtaining a first speech feature signal according to the first speed-adjusted speech signal; obtaining a detection result according to a first keyword detection result corresponding to the first speech feature signal, the detection result indicating whether a target keyword exists in the to-be-detected speech signal; and performing an operation corresponding to the target keyword in response to determining that the target keyword exists according to the detection result.

Abstract: Automatic speech recognition techniques are implemented in resource constrained devices such as edge devices in internet of things where on-device speech recognition is required for low latency and privacy preservation. Existing neural network models for speech recognition have a large size and are not suitable for deployment in such devices. The present disclosure provides an architecture of a size constrained neural network and a method of training the size constrained neural network. The architecture of the size constrained neural network provides a way of increasing or decreasing number of feature blocks to achieve an accuracy-model size trade off. The method of training the size constrained neural network comprises creating a training dataset with short utterances and training the size constrained neural network with the training dataset to learn short term dependencies in the utterances. The trained size constrained neural network model is suitable for deployment in resource constrained devices.

Abstract: A disclosed method includes monitoring an audio signal energy level while having a noise suppressor deactivated to conserve battery power, buffering the audio signal in response to a detected increase in the audio energy level, activating and running a voice activity detector on the audio signal in response to the detected increase in the audio energy level and activating and running a noise estimator in response to voice being detected in the audio signal by the voice activity detector. The method may further include activating and running the noise suppressor only if the noise estimator determines that noise suppression is required. The method activates and runs a noise type classifier to determine the noise type based on information received from the noise estimator and selects a noise suppressor algorithm, from a group of available noise suppressor algorithms, where the selected noise suppressor algorithm is the most power consumption efficient.

Abstract: This disclosure relates generally to methods and systems for building an intelligent analytical platform to enable a device fabrication in material science. Material engineers and design engineers may face various challenges with existing knowledge, as more time and efforts are required in finding a relevant knowledge from the existing knowledge, mainly due to the unstructured form, for fabricating new devices. The present disclosure solves the technical problem of finding the relevant knowledge out of the existing knowledge, in a structured form by building an analytical platform. The unstructured format of the existing knowledge of the fabrication process is transformed into a structured format in terms of operation sequence knowledge graphs, using a set of artificial intelligence (AI) and machine learning models, and a knowledge representation model of the fabrication process. The structured format of the existing knowledge is hierarchically arranged to build the analytical platform.

Abstract: The present approaches are generally related to an agent automation framework that is capable of extracting meaning from user utterances, such as requests received by a virtual agent (e.g., a chat agent), and suitably responding to these user utterances. In certain aspects, the agent automation framework includes a NLU framework and an intent-entity model having defined intents and entities that are associated with sample utterances. The NLU framework may include a meaning extraction subsystem designed to generate meaning representations for the sample utterances of the intent-entity model to construct an understanding model, as well as generate meaning representations for a received user utterance to construct an utterance meaning model. The disclosed NLU framework may include a meaning search subsystem that is designed to search the meaning representations of the understanding model to locate matches for meaning representations of the utterance meaning model.

Abstract: Provided is a sound quality detection method, including: acquiring a plurality of audio files to be detected, wherein the plurality of audio files are homologous audio files; acquiring at least one audio feature of each of the plurality of audio files by performing feature extraction on the audio file, and generating a correspondence list between the at least one audio feature of each of the plurality of audio files and an audio file identifier; and determining, using a sound quality detection model, a sound quality score of each of the plurality of audio files based on the correspondence list between the at least one audio feature of each of the plurality of audio files and the audio file identifier, wherein the sound quality detection model is configured to detect sound quality of homologous audio files.

Abstract: In the present invention, a method for searching multilingual keywords in mixlingual speech corpus is proposed. This method is capable of searching audio as well as text keywords. The capability of audio search enables it to search out-of-vocabulary (OOV) words. The capability of searching text keywords enables it to perform semantic search. An advanced application of searching keyword translations in mixlingual speech corpus is also possible within posteriorgram framework with this system. Also, a technique for combining information from text and audio keywords is given which further enhances the search performance. This system is based on multiple posteriorgrams based on articulatory classes trained with multiple languages.

Abstract: Aspects of the disclosure include computer-implemented methods and systems for providing generative adversarial network (GAN) digital image data. GAN digital image data corresponding to a suggested transaction for an identified customer can be determined.

Abstract: Methods and systems for processing user input to a computing system are disclosed. The computing system has access to an audio input and a visual input such as a camera. Face detection is performed on an image from the visual input, and if a face is detected this triggers the recording of audio and making the audio available to a speech processing function. Further verification steps can be combined with the face detection step for a multi-factor verification of user intent to interact with the system.