International Business Machines Corporation

Nuance Communications, Inc.

Sumitomo Electric Industries, Limited

Teikoku Printing Inks Mfg. Co., Ltd.

Osamu Ichikawa

Takashi Fukuda

Masafumi Nishimura

Bhuvana Ramabhadran

Gakuto Kurata

Masayuki A Suzuki

Yasunori Saito

Samuel Jonathan Thomas

Tetsuya Takiguchi

Futoshi Iwama

Steven John Rennie

Toshio Oshima

Eisuke Sasaoka

Tohru Nagano

A student neural network may be trained by a computer-implemented method, including: inputting common input data to each teacher neural network among a plurality of teacher neural networks to obtain a soft label output among a plurality of soft label outputs from each teacher neural network among the plurality of teacher neural networks, and training a student neural network with the input data and the plurality of soft label outputs.

Training of student neural network with teacher neural networks

A student neural network may be trained by a computer-implemented method, including: selecting a teacher neural network among a plurality of teacher neural networks, inputting an input data to the selected teacher neural network to obtain a soft label output generated by the selected teacher neural network, and training a student neural network with at least the input data and the soft label output from the selected teacher neural network.

Training of student neural network with switched teacher neural networks

A technique for constructing a model supporting a plurality of domains is disclosed. In the technique, a plurality of teacher models, each of which is specialized for different one of the plurality of the domains, is prepared. A plurality of training data collections, each of which is collected for different one of the plurality of the domains, is obtained. A plurality of soft label sets is generated by inputting each training data in the plurality of the training data collections into corresponding one of the plurality of the teacher models. A student model is trained using the plurality of the soft label sets.

Teacher and student learning for constructing mixed-domain model

A technique for training a neural network including an input layer, one or more hidden layers and an output layer, in which the trained neural network can be used to perform a task such as speech recognition. In the technique, a base of the neural network having at least a pre-trained hidden layer is prepared. A parameter set associated with one pre-trained hidden layer in the neural network is decomposed into a plurality of new parameter sets. The number of hidden layers in the neural network is increased by using the plurality of the new parameter sets. Pre-training for the neural network is performed.

Pre-training of neural network by parameter decomposition

A method is provided for training a neural network-based (NN-based) acoustic model. The method includes receiving, by a processor, the neural network-based (NN-based) acoustic model, trained by a one-hot scheme and having an input layer, a set of middle layers, and an original output layer. At least each of the middle layers subsequent to a first one of the middle layers have trained parameters. The method further includes stacking, by the processor, a new output layer on the original output layer of the NN-based acoustic model to form a new NN-based acoustic model. The new output layer has a same size as the original output layer. The method also includes retraining, by the processor, only the new output layer and the original output layer of the new NN-based acoustic model in the one-hot scheme, with the trained parameters of middle layers subsequent to at least the first one being fixed.

Neural network-based acoustic model with softening target-layer

A computer-implemented method is provided. The computer-implemented method is performed by a speech recognition system having at least a processor. The method further includes performing a speech recognition operation on the audio signal data to decode the audio signal data into a textual representation based on the estimated sound identification information from a neural network having periodic indications and components of a frequency spectrum of the audio signal data inputted thereto. The neural network includes a plurality of fully-connected network layers having a first layer that includes a plurality of first nodes and a plurality of second nodes. The method further comprises training the neural network by initially isolating the periodic indications from the components of the frequency spectrum in the first layer by setting weights between the first nodes and a plurality of input nodes corresponding to the periodic indications to 0.

Sound identification utilizing periodic indications

Acoustic change is detected by a method including preparing a first Gaussian Mixture Model (GMM) trained with first audio data of first speech sound from a speaker at a first distance from an audio interface and a second GMM generated from the first GMM using second audio data of second speech sound from the speaker at a second distance from the audio interface; calculating a first output of the first GMM and a second output of the second GMM by inputting obtained third audio data into the first GMM and the second GMM; and transmitting a notification in response to determining at least that a difference between the first output and the second output exceeds a threshold. Each Gaussian distribution of the second GMM has a mean obtained by shifting a mean of a corresponding Gaussian distribution of the first GMM by a common channel bias.

Acoustic change detection for robust automatic speech recognition based on a variance between distance dependent GMM models

A method, computer system, and a computer program product for generating a plurality of voice data having a particular speaking style is provided. The present invention may include preparing a plurality of original voice data corresponding to at least one word or at least one phrase is prepared. The present invention may also include attenuating a low frequency component and a high frequency component in the prepared plurality of original voice data. The present invention may then include reducing power at a beginning and an end of the prepared plurality of original voice data. The present invention may further include storing a plurality of resultant voice data obtained after the attenuating and the reducing.

Generation of voice data as data augmentation for acoustic model training

A method for learning a neural network having a plurality of filters for extracting local features performed by a computing device is disclosed. The computing device calculates a plurality of projection parameter sets by analyzing one or more training data. The plurality of the projection parameter sets define a projection of each training data into a new space and each projection parameter set has a same size as the filters in the neural network. At least part of the plurality of the projection parameter sets is set as initial parameters of at least part of the plurality of the filters in the neural network for training.

Learning of neural network

A computer-implemented method for processing a speech signal, includes: identifying speech segments in an input speech signal; calculating an upper variance and a lower variance, the upper variance being a variance of upper spectra larger than a criteria among speech spectra corresponding to frames in the speech segments, the lower variance being a variance of lower spectra smaller than a criteria among the speech spectra corresponding to the frames in the speech segments; determining whether the input speech signal is a special input speech signal using a difference between the upper variance and the lower variance; and performing speech recognition of the input speech signal which has been determined to be the special input speech signal, using a special acoustic model for the special input speech signal.

Processing of speech signal

A computer-implemented method is provided. The computer-implemented method is performed by a speech recognition system having at least a processor. The method includes estimating sound identification information from a neural network having periodic indications and components of a frequency spectrum of an audio signal data inputted thereto. The method further includes performing a speech recognition operation on the audio signal data to decode the audio signal data into a textual representation based on the estimated sound identification information. The neural network includes a plurality of fully-connected network layers having a first layer that includes a plurality of first nodes and a plurality of second nodes. The method further comprises training the neural network by initially isolating the periodic indications from the components of the frequency spectrum in the first layer by setting weights between the first nodes and a plurality of input nodes corresponding to the periodic indications to 0.

A method, for testing words defined in a pronunciation lexicon used in an automatic speech recognition (ASR) system, is provided. The method includes: obtaining test sentences which can be accepted by a language model used in the ASR system. The test sentences cover words defined in the pronunciation lexicon. The method further includes obtaining variations of speech data corresponding to each test sentence, and obtaining a plurality of texts by recognizing the variations of speech data, or a plurality of texts generated by recognizing the variation of speech data. The method also includes constructing a word graph, using the plurality of texts, for each test sentence, where each word in the word graph corresponds to each word defined in the pronunciation lexicon; and determining whether or not all or parts of words in a test sentence are present in a path of the word graph derived from the test sentence.

Testing words in a pronunciation lexicon

A method, a system, and a computer program product detect a clipping event in audio signals. The method includes digitalizing audio signals having limited frequency bands, at a sampling frequency which is greater than two times as large as the maximum frequency component of the audio signal; and detecting a clipping event of the audio signals, based on magnitudes of spectrum in a bandwidth which is greater than or equal to the limited frequency band. The sampling frequency may be greater than or equal to three times as large as the maximum frequency component of the audio signal. The detection of a clipping event may include determining, for each frame, whether or not a sum or average of the magnitudes of spectrum at the bandwidth which is greater than or equal to the limited frequency band is larger than a predetermined threshold.

Detection of clipping event in audio signals

A method and system for generating training data for a target domain using speech data of a source domain. The training data generation method including: reading out a Gaussian mixture model (GMM) of a target domain trained with a clean speech data set of the target domain; mapping, by referring to the GMM of the target domain, a set of source domain speech data received as an input to the set of target domain speech data on a basis of a channel characteristic of the target domain speech data; and adding a noise of the target domain to the mapped set of source domain speech data to output a set of pseudo target domain speech data.

Growing community of inventors

Yokohama, Japan

Osamu Ichikawa