Nippon Telegraph and Telephone Corporation

Takuya Yoshioka

Tomohiro Nakatani

Keisuke Kinoshita

Masato Miyoshi

Marc Delcroix

Takuya Higuchi

Shoko Araki

Atsunori Ogawa

Nobutaka Ito

Takafumi Hikichi

Mehrez Souden

To begin with, an acoustic model training apparatus extracts speech features representing speech characteristics, and calculates an acoustic-condition feature representing a feature of an acoustic condition of the speech data using an acoustic-condition calculation model that is represented as a neural network, based on an acoustic-condition calculation model parameter characterizing the acoustic-condition calculation model. The acoustic model training apparatus then generates an adjusted parameter that is an acoustic model parameter adjusted based on the acoustic-condition feature, the acoustic model parameter characterizing an acoustic model represented as a neural network to which an output layer of the acoustic-condition calculation model is coupled. The acoustic model training apparatus then updates the acoustic model parameter based on the adjusted parameter and the speech features, and updates the acoustic-condition calculation model parameters based on the adjusted parameter and the speech features.

Acoustic model training method, speech recognition method, acoustic model training apparatus, speech recognition apparatus, acoustic model training program, and speech recognition program

An observation feature value vector is calculated based on observation signals recorded at different positions in a situation in which target sound sources and background noise are present in a mixed manner; masks associated with the target sound sources and a mask associated with the background noise are estimated; a spatial correlation matrix of the target sound sources that includes the background noise is calculated based on the masks associated with the observation signals and the target sound sources; a spatial correlation matrix of the background noise is calculated based on the masks associated with the observation signals and the background noise; and a spatial correlation matrix of the target sound sources is estimated based on the matrix obtained by weighting each of the spatial correlation matrices by predetermined coefficients.

Spatial correlation matrix estimation device, spatial correlation matrix estimation method, and spatial correlation matrix estimation program

A noise estimation apparatus which estimates a non-stationary noise component on the basis of the likelihood maximization criterion is provided. The noise estimation apparatus obtains the variance of a noise signal that causes a large value to be obtained by weighted addition of the sums each of which is obtained by adding the product of the log likelihood of a model of an observed signal expressed by a Gaussian distribution in a speech segment and a speech posterior probability in each frame, and the product of the log likelihood of a model of an observed signal expressed by a Gaussian distribution in a non-speech segment and a non-speech posterior probability in each frame, by using complex spectra of a plurality of observed signals up to the current frame.

Noise estimation apparatus, noise estimation method, noise estimation program, and recording medium

The initial values of parameter estimates are set, including reverberation parameter estimates, which includes a regression coefficient used in a linear convolutional operation for calculating an estimated value of reverberation included in an observed signal, source parameter estimates, which includes estimated values of a linear prediction coefficient and a prediction residual power that identify the power spectrum of a source signal, and noise parameter estimates, which include noise power spectrum estimates. Then, the maximum likelihood estimation is used to alternately repeat processing for updating at least one of the reverberation parameter estimates and the noise parameter estimates and processing for updating the source parameter estimates until a predetermined termination condition is satisfied.

Signal enhancement device, method thereof, program, and recording medium

Provided is a signal distortion elimination apparatus comprising: an inverse filter application means that outputs the signal obtained by applying an inverse filter to an observed signal as a restored signal when a predetermined iteration termination condition is met and outputs the signal obtained by applying the inverse filter to the observed signal as an ad-hoc signal when the predetermined iteration termination condition is not met; a prediction error filter calculation means that segments the ad-hoc signal into frames and outputs a prediction error filter of each frame obtained by performing linear prediction analysis of the ad-hoc signal of each frame; an inverse filter calculation means that calculates an inverse filter such that a concatenation of innovation estimates of the respective frames becomes mutually independent among their samples, where the innovation estimate of a single frame (an innovation estimate) is the signal obtained by applying the prediction error filter of the corresponding frame to the ad-hoc signal of the corresponding frame, and outputs the inverse filter; and a control means that iteratively executes the inverse filter application means, the prediction error filter calculation means and the inverse filter calculation means until the iteration termination condition is met.

Signal distortion elimination apparatus, method, program, and recording medium having the program recorded thereon

A sound source model storage section stores a sound source model that represents an audio signal emitted from a sound source in the form of a probability density function. An observation signal, which is obtained by collecting the audio signal, is converted into a plurality of frequency-specific observation signals each corresponding to one of a plurality of frequency bands. Then, a dereverberation filter corresponding to each frequency band is estimated by using the frequency-specific observation signal for the frequency band on the basis of the sound source model and a reverberation model that represents a relationship for each frequency band among the audio signal, the observation signal and the dereverberation filter. A frequency-specific target signal corresponding to each frequency band is determined by applying the dereverberation filter for the frequency band to the frequency-specific observation signal for the frequency band, and the resulting frequency-specific target signals are integrated.

Growing community of inventors

Kyoto, Japan

Takuya Yoshioka