Mitsubishi Electric Research Laboratories, Inc.

Google Inc.

John R Hershey

Jonathan Le Roux

Shinji Watanabe

Takaaki Hori

Bret Harsham

Chiori Hori

Daniel Nikolaev Nikovski

Felix Weninger

Petros T Boufounos

Tim Marks

Gordon P Wichern

Aren Jansen

Zhuo Chen

Scott Wisdom

Hiroshi Seki

Efthymios Tzinis

William Yerazunis

Yuichi Taguchi

Dirk Brinkman

Laurent Daudet

Hakan Erdogan

Alan W Esenther

Vamsi K Potluru

Umut Simsekli

Tsubasa Ochiai

Creighton K Heakulani

Kang Kang

Meng Yu

Shane Settle

Zhongqiu Wang

Apparatus and methods related to separation of audio sources are provided. The method includes receiving an audio waveform associated with a plurality of video frames. The method includes estimating, by a neural network, one or more audio sources associated with the plurality of video frames. The method includes generating, by the neural network, one or more audio embeddings corresponding to the one or more estimated audio sources. The method includes determining, based on the audio embeddings and a video embedding, whether one or more audio sources of the one or more estimated audio sources correspond to objects in the plurality of video frames. The method includes predicting, by the neural network and based on the one or more audio embeddings and the video embedding, a version of the audio waveform comprising audio sources that correspond to objects in the plurality of video frames.

Audio-visual separation of on-screen sounds based on machine learning models

A speech recognition system includes a plurality of microphones to receive acoustic signals including speech signals, an input interface to generate multichannel inputs from the acoustic signals, one or more storages to store a multichannel speech recognition network, wherein the multichannel speech recognition network comprises mask estimation networks to generate time-frequency masks from the multichannel inputs, a beamformer network trained to select a reference channel input from the multichannel inputs using the time-frequency masks and generate an enhanced speech dataset based on the reference channel input and an encoder-decoder network trained to transform the enhanced speech dataset into a text. The system further includes one or more processors, using the multichannel speech recognition network in association with the one or more storages, to generate the text from the multichannel inputs, and an output interface to render the text.

System and method for multichannel end-to-end speech recognition

A method for performing multi-label classification includes extracting a feature vector from an input vector including input data by a feature extractor, determining, by a label predictor, a relevant vector including relevant labels having relevant scores based on the feature vector, updating a binary masking vector by masking pre-selected labels having been selected in previous label selections, applying the updated binary masking vector to the relevant vector such that the relevant label vector is updated to exclude the pre-selected labels from the relevant labels, and selecting a relevant label from the updated relevant label vector based on the relevant scores of the updated relevant label vector.

Method and system for multi-label classification

Systems and methods for a speech recognition system for recognizing speech including overlapping speech by multiple speakers. The system including a hardware processor. A computer storage memory to store data along with having computer-executable instructions stored thereon that, when executed by the processor is to implement a stored speech recognition network. An input interface to receive an acoustic signal, the received acoustic signal including a mixture of speech signals by multiple speakers, wherein the multiple speakers include target speakers. An encoder network and a decoder network of the stored speech recognition network are trained to transform the received acoustic signal into a text for each target speaker. Such that the encoder network outputs a set of recognition encodings, and the decoder network uses the set of recognition encodings to output the text for each target speaker. An output interface to transmit the text for each target speaker.

Methods and systems for recognizing simultaneous speech by multiple speakers

Systems and methods for audio signal processing including an input interface to receive a noisy audio signal including a mixture of target audio signal and noise. An encoder to map each time-frequency bin of the noisy audio signal to one or more phase-related value from one or more phase quantization codebook of phase-related values indicative of the phase of the target signal. Calculate, for each time-frequency bin of the noisy audio signal, a magnitude ratio value indicative of a ratio of a magnitude of the target audio signal to a magnitude of the noisy audio signal. A filter to cancel the noise from the noisy audio signal based on the phase-related values and the magnitude ratio values to produce an enhanced audio signal. An output interface to output the enhanced audio signal.

Methods and systems for enhancing audio signals corrupted by noise

A speech recognition system includes an input device to receive voice sounds, one or more processors, and one or more storage devices storing parameters and program modules including instructions which cause the one or more processors to perform operations. The operations include extracting an acoustic feature sequence from audio waveform data converted from the voice sounds, encoding the acoustic feature sequence into a hidden vector sequence using an encoder network having encoder network parameters, predicting first output label sequence probabilities by feeding the hidden vector sequence to a decoder network having decoder network parameters, predicting second output level sequence probabilities by a hybrid network using character-base language models (LMs) and word-level LMs; and searching, using a label sequence search module, for an output label sequence having a highest sequence probability by combining the first and second output label sequence probabilities provided from the decoder network and the hybrid network.

Method and apparatus for open-vocabulary end-to-end speech recognition

A method for training a multi-language speech recognition network includes providing utterance datasets corresponding to predetermined languages, inserting language identification (ID) labels into the utterance datasets, wherein each of the utterance datasets is labelled by each of the language ID labels, concatenating the labeled utterance datasets, generating initial network parameters from the utterance datasets, selecting the initial network parameters according to a predetermined sequence, and training, iteratively, an end-to-end network with a series of the selected initial network parameters and the concatenated labeled utterance datasets until a training result reaches a threshold.

Method and apparatus for multi-lingual end-to-end speech recognition

A method for transforms input signals, by first defining a model for transforming the input signals, wherein the model is specified by constraints and a set of model parameters. An iterative inference procedure is derived from the model and the set of model parameters and unfolded into a set of layers, wherein there is one layer for each iteration of the procedure, and wherein a same set of network parameters is used by all layers. A neural network is formed by untying the set of network parameters such that there is one set of network parameters for each layer and each set of network parameters is separately maintainable and separately applicable to the corresponding layer. The neural network is trained to obtain a trained neural network, and then input signals are transformed using the trained neural network to obtain output signals.

Neural networks for transforming signals

Systems and methods for an audio signal processing system for transforming an input audio signal. A processor implements steps of a module by inputting an input audio signal into a spectrogram estimator to extract an audio feature sequence, and process the audio feature sequence to output a set of estimated spectrograms. Processing the set of estimated spectrograms and the audio feature sequence using a spectrogram refinement module, to output a set of refined spectrograms. Wherein the processing of the spectrogram refinement module is based on an iterative reconstruction algorithm. Processing the set of refined spectrograms for the one or more target audio signals using a signal refinement module, to obtain the target audio signal estimates. An output interface to output the optimized target audio signal estimates. Wherein the module is optimized by minimizing an error using an optimizer stored in the memory.

Methods and systems for end-to-end speech separation with unfolded iterative phase reconstruction

A system for generating a word sequence includes one or more processors in connection with a memory and one or more storage devices storing instructions causing operations that include receiving first and second input vectors, extracting first and second feature vectors, estimating a first set of weights and a second set of weights, calculating a first content vector from the first set of weights and the first feature vectors, and calculating a second content vector, transforming the first content vector into a first modal content vector having a predetermined dimension and transforming the second content vector into a second modal content vector having the predetermined dimension, estimating a set of modal attention weights, generating a weighted content vector having the predetermined dimension from the set of modal attention weights and the first and second modal content vectors, and generating a predicted word using the sequence generator.

Method and system for multi-modal fusion model

A method and for training a language model to reduce recognition errors, wherein the language model is a recurrent neural network language model (RNNLM) by first acquiring training samples. An automatic speech recognition system (ASR) is applied to the training samples to produce recognized words and probabilites of the recognized words, and an N-best list is selected from the recognized words based on the probabilities. determining word errors using reference data for hypotheses in the N-best list. The hypotheses are rescored using the RNNLM. Then, we determine gradients for the hypotheses using the word errors and gradients for words in the hypotheses. Lastly, parameters of the RNNLM are updated using a sum of the gradients.

Method and system for training language models to reduce recognition errors

A method transforms a noisy audio signal to an enhanced audio signal, by first acquiring the noisy audio signal from an environment. The noisy audio signal is processed by an enhancement network having network parameters to jointly produce a magnitude mask and a phase estimate. Then, the magnitude mask and the phase estimate are used to obtain the enhanced audio signal.

Method for enhancing audio signal using phase information

A method and system processes utterances that are acquired either from an automatic speech recognition (ASR) system or text. The utterances have associated identities of each party, such as role A utterances and role B utterances. The information corresponding to utterances, such as word sequence and identity, are converted to features. Each feature is received in an input layer of a neural network (NN). A dimensionality of each feature is reduced, in a projection layer of the NN, to produce a reduced dimensional feature. The reduced dimensional feature is processed to provide probabilities of labels for the utterances.

Method and system for role dependent context sensitive spoken and textual language understanding with neural networks

A method for processing a voice command using a statistical dialog model determines a belief state as a probability distribution over states organized in a hierarchy with a parent-child relationship of nodes representing the states. The belief state includes the hierarchy of state variables defining probabilities of each state to correspond to the voice command and a probability of a state of a child node in the hierarchy is conditioned on a probability of a state of a corresponding parent node. A system action is selected based on the belief state.

Growing community of inventors

Winchester, MA, United States of America

John R Hershey