Multichannel microphone-based reverberation time estimation method and device which use deep neural network

Patent metadata

Field	Value
Publication number	US-10854218-B2
Application number	US-201716469938-A
Country	US
Kind code	B2
Filing date	Dec 15, 2017
Priority date	Dec 15, 2016
Publication date	Dec 1, 2020
Grant date	Dec 1, 2020

Abstract

A multichannel microphone-based reverberation time estimation method and device which use a deep neural network (DNN) are disclosed. A multichannel microphone-based reverberation time estimation method using a DNN, according to one embodiment, comprises the steps of: receiving a voice signal through a multichannel microphone; deriving a feature vector including spatial information by using the inputted voice signal; and estimating the degree of reverberation by applying the feature vector to the DNN.

First claim

What is claimed is:

1. A method for multichannel microphone-based reverberation time estimation using a deep neural network (DNN), the method comprising:
receiving an input of a voice signal through a multichannel microphone;
deriving a feature vector that includes spatial information using the input voice signal, and
estimating a degree of reverberation by applying the feature vector to the DNN,
wherein the deriving of the feature vector comprises:
deriving a negative-side variance (NSV) by deriving time and frequency information from the input voice signal using a short-time Fourier transform (STFT) and by deriving distribution of envelopes for each frequency band based on the derived time and frequency information; and
deriving a cross-correlation function representing a correlation between two microphones in the input voice signal, and

wherein the estimating of the degree of reverberation comprises estimating a reverberation time by using the derived NSV and the cross-correlation function as an input of the DNN.

2. The method of claim 1 , wherein the receiving of the voice signal through the multichannel microphone comprises estimating relative spatial information between voice signals input using the multichannel microphone.

3. The method of claim 1 , wherein the deriving of the NSV comprises:
deriving a log-energy envelope from a domain of the STFT;
deriving a gradient from the log-energy envelope using a least squares linear fitting; and
deriving an NSV for estimating a reverberation time having a negative gradient, excluding a reverberation time having a positive gradient.

4. The method of claim 1 , wherein the estimating of the degree of reverberation by applying the feature vector to the DNN comprises estimating a reverberation time by using the derived NSV as an input of the DNN.

5. The method of claim 1 , wherein the DNN comprises three hidden layers, and each of the hidden layers is configured to be finely adjusted through a pre-training process using a plurality of epochs.

6. An apparatus for multichannel microphone-based reverberation time estimation using a deep neural network (DNN), the apparatus comprising:
an inputter configured to receive an input of a voice signal through a multichannel microphone;
a feature vector extractor configured to derive a feature vector that includes spatial information using the input voice signal; and
a reverberation estimator configured to estimate a degree of reverberation by applying the feature vector to the DNN,
wherein the feature vector extractor comprises:
a negative-side variance (NSV) deriver configured to derive an NSV by deriving time and frequency information from the input voice signal using a short-time Fourier transform (STFT) and by deriving distribution of envelopes for each frequency band based on the derived time and frequency information; and
a cross-correlation function deriver configured to derive a cross-correlation function representing a correlation between two microphones in the input voice signal, and

wherein the reverberation estimator is configured to estimate a reverberation time by using the derived NSV the cross-correlation function as an input of the DNN.

7. The apparatus of claim 6 , wherein the NSV deriver is configured to derive a log-energy envelope from a domain of the STFT, to derive a gradient from the log-energy envelope using a least squares linear fitting, and to derive an NSV for estimating a reverberation having a negative gradient, excluding a reverberation time having a positive gradient.

8. The apparatus of claim 6 , wherein the reverberation estimator is configured to estimate a reverberation time by using the derived NSV as an input of the DNN.

9. The apparatus of claim 6 , wherein the DNN comprises three hidden layers, and each of the hidden layers is configured to be finely adjusted through a pre-training process using a plurality of epochs.

Assignees

• Univ Hanyang Ind Univ Coop Found

Inventors

• Chang Joon-Hyuk
• Lee Myung In

Classifications

• G06N3/09

  Supervised learning · CPC title

• G06N3/0499

  Feedforward networks · CPC title

• G10L2021/02165

  Two microphones, one receiving mainly the noise signal and the other one mainly the speech signal · CPC title

• G10L2021/02082

  the noise being echo, reverberation of the speech · CPC title

• G10L21/0216 Primary

  characterised by the method used for estimating noise · CPC title