Nippon Telegraph and Telephone Corporation

Marc Delcroix

Tomohiro Nakatani

Keisuke Kinoshita

Atsunori Ogawa

Shoko Araki

Shigeki Karita

Tsubasa Ochiai

Takuya Yoshioka

Hiroshi Sato

Hiroaki Ito

Takuya Higuchi

Yuma Koizumi

Nobutaka Ito

Yuki Kubo

Michael Hentschel

Mehrez Souden

Naohiro Tawara

Lukas Drude

Thilo Christoph Von Neumann

A data processing device includes processing circuitry configured to extract a second word corresponding to a first word included in first text from among a plurality of words belonging to a predetermined domain, repeat processing in the extraction for all words included in the first text to generate a confusion network that expresses a plurality of sentence possibilities with one network configuration and is an expression format of a word sequence, and search for a grammatically correct word string in the confusion network using a language model that evaluates grammatical correctness of the word string, and select a word string to be output.

Data processing device, data processing method, and data processing program

An audio signal processing apparatus () includes a first auxiliary feature conversion unit () and a second auxiliary feature conversion unit () that convert a plurality of signals relating to processing of an audio signal of a target speaker into a plurality of auxiliary features for the plurality of signals using a plurality of auxiliary neural networks corresponding to the plurality of signals, and an audio signal processing unit () that estimates information regarding an audio signal of the target speaker included in a mixed audio signal using a main neural network based on an input feature of the mixed audio signal and the plurality of auxiliary features, wherein the plurality of signals relating to processing of the audio signal of the target speaker are two or more pieces of information of different modalities.

Speech signal processing device, speech signal processing method, speech signal processing program, training device, training method, and training program

A signal processing device includes processing circuitry configured to receive an input of extraction target information indicating which audio class of an audio signal is to be extracted from a mixture audio signal constituted by a mixture of audio signals of a plurality of audio classes, and output a result of extracting the audio signal of the audio class indicated by the extraction target information from the mixture audio signal, with a neural network by using a feature value of the mixture audio signal and the extraction target information.

Signal processing device, signal processing method, signal processing program, learning device, learning method, and learning program

Disclosed is a model adaptation technology of a language model with higher adaptability. An aspect of the present disclosure relates to an apparatus includes a first neural network unit that transforms an input symbol and outputs an intermediate state; and a second neural network unit that transforms input auxiliary information and the intermediate state and predicts a symbol following the input symbol, wherein the second neural network unit includes a plurality of hidden layers receiving, as input, the intermediate state and auxiliary information, and pieces of the auxiliary information input to each hidden layer are different from each other.

Apparatus, method, and program for utilizing language model

A mask estimation apparatus includes processing circuitry configured to estimate, for a target segment to be processed among a plurality of segments of a continuous time, a first mask which is an occupancy ratio of a target signal to an observation signal of the target segment, based on a first feature obtained from a plurality of the observation signals of the target segment recorded at a plurality of locations, and estimate a parameter for modeling a second feature and a second mask which is an occupancy ratio of the target signal to the observation signal based on an estimation result of the first mask in the target segment and the second feature obtained from the plurality of the observation signals of the target segment.

Mask estimation device, mask estimation method, and mask estimation program

Provided is a target sound extraction technique based on a steering vector generation method enabling instability in a calculation to be prevented when a neural network is trained by using an error back propagation method to reduce an estimation error of a beamformer. A target sound signal generation apparatus generates a target sound signal ycorresponding to a target sound included in an observed sound from an observed signal vector xcorresponding to the observed sound collected by using a plurality of microphones. The target sound signal generation apparatus includes a mask generation unit, a steering vector generation unit, a beamformer vector generation unit, and a target sound signal generation unit. The mask generation unit is configured as a neural network trained by using an error back propagation method. The steering vector generation unit generates a steering vector hby determining an eigenvector corresponding to a maximum eigenvalue of a predetermined matrix generated from the observed signal vector xand a mask γby using a power method.

Target sound signal generation apparatus, target sound signal generation method, and program

A learning device includes a memory, and processing circuitry coupled to the memory and configured to receive an input of a plurality of series for learning having known accuracy, and learn a model represented by a neural network, the model being capable of determining accuracy levels of two series when given feature amounts of the two series among the plurality of series.

Learning device, learning method, and learning program

A reranking device include a hypothesis input unit configured to receive input of N-best hypotheses associated with scores of a speech recognition accuracy; a hypothesis selection unit configured to select two hypotheses to be determined from among the input N-best hypotheses. Further, there is a determination unit configured to determine which accuracy of two hypotheses is higher by using: a plurality of first auxiliary model to M-th auxiliary model represented by such a neural network as to be capable of converting, when the selected two hypotheses are given, the two hypotheses into hidden state vectors, and determining which of the two hypotheses is higher based on the hidden state vectors of the two hypotheses; and a main model represented by such a neural network as to be capable of determining which of the two hypotheses is higher based on the hidden state vectors of the two hypotheses.

Determination device, training device, determination method, and determination program

A signal processing device according to an embodiment of the present invention includes: a conversion unit configured to convert an input mixed acoustic signal into a plurality of first internal states, a weighting unit configured to generate a second internal state which is a weighted sum of the plurality of first internal states based on auxiliary information regarding an acoustic signal of a target sound source when the auxiliary information is input, and generate the second internal state by selecting one of the plurality of first internal states when the auxiliary information is not input, and a mask estimation unit configured to estimate a mask based on the second internal state.

Signal processing apparatus, learning apparatus, signal processing method, learning method and program

A determination device includes a memory, and processing circuitry coupled to the memory and configured to accept input of a plurality of sequences provided as candidates for a solution to one given input, and determine, for two sequences of the plurality of sequences, a sequence that has a higher accuracy than the other sequence of the two sequences, using a model expressed as a neural network.

Determination device, determination method, and determination program

Features are extracted from an observed speech signal including at least speech of multiple speakers including a target speaker. A mask is calculated for extracting speech of the target speaker based on the features of the observed speech signal and a speech signal of the target speaker serving as adaptation data of the target speaker. The signal of the speech of the target speaker is calculated from the observed speech signal based on the mask. Speech of the target speaker can be extracted from observed speech that includes speech of multiple speakers.

Mask calculation device, cluster weight learning device, mask calculation neural network learning device, mask calculation method, cluster weight learning method, and mask calculation neural network learning method

A time-variant noise spatial covariance matrix is estimated effectively. Using time-frequency-divided observation signals based on observation signals acquired by collecting acoustic signals emitted from one or a plurality of sound sources and mask information expressing the occupancy probability of a component of each of the time-frequency-divided observation signals that corresponds to each noise source, a time-independent first noise spatial covariance matrix corresponding to the time-frequency-divided observation signals and the mask information belonging to a long time interval is acquired for each noise source. Further, using the mask information of each of a plurality of different short time intervals, a mixture weight corresponding to each noise source in each short time interval is acquired. Furthermore, a time-variant third noise spatial covariance matrix is acquired, the third noise spatial covariance matrix being based on a time-variant second noise spatial covariance matrix, which corresponds to the time-frequency-divided observation signals and the mask information belonging to each short time interval and relates to noise formed by adding together all of the noise sources, and a weighted sum of the first noise spatial covariance matrices with the mixture weights of the respective short time intervals.

Noise spatial covariance matrix estimation apparatus, noise spatial covariance matrix estimation method, and program

A learning device () includes a feature extracting unit () that extracts features of speech from speech data for training, a probability calculating unit () that, on the basis of the features of speech, performs prefix searching using a speech recognition model of which a neural network is representative, and calculates a posterior probability of a recognition character string to obtain a plurality of hypothetical character strings, an error calculating unit () that calculates an error by word error rates of the plurality of hypothetical character strings and a correct character string for training, and obtains a parameter for the entire model that minimizes an expected value of summation of loss in the word error rates, and an updating unit () that updates a parameter of the model in accordance with the parameter obtained by the error calculating unit ().

Learning device and method for updating a parameter of a speech recognition model

An estimation device includes a memory, and processing circuitry coupled to the memory and configured to receive an input of an input audio signal that is an audio signal in which sounds from a plurality of sound sources are mixed, and an input of supplemental information, and output an estimation result of mask information that identifies a mask for extracting a sound of any one of the sound sources included in an entire or a part of a signal included in the input audio signal, the signal being identified by the supplemental information, cause a neural network to iterate a process of outputting the estimation result of the mask information, and cause the neural network to output an estimation result of the mask information for a different sound source, by inputting a different piece of the supplemental information to the neural network at each iteration.

Estimation device, learning device, estimation method, learning method, and recording medium

A signal processing device includes a power estimating unit that treats the feature quantity of a signal including reverberation as the input; inputs an observation feature quantity corresponding to an observation signal to a neural network which is learnt in such a way that the estimate value of the feature quantity corresponding to the power of the signal having reduced reverberation, from among the input signal, is output; and estimates the estimate value of the feature quantity corresponding to the power of the signal having reduced reverberation and corresponding to the observation signal. Moreover, the signal processing device includes a regression coefficient estimating unit that uses the estimate value of the feature quantity corresponding to the power as obtained as the estimation result by the power estimating unit, and estimates a regression coefficient of the autoregressive process for generating the observation signal.

Growing community of inventors

Kyoto, Japan

Marc Delcroix