Speech recognition apparatus and method in noisy circumstances

• US 5,749,068 A
• Filed: 10/17/1996
• Issued: 05/05/1998
• Est. Priority Date: 03/25/1996
• Status: Expired due to Term

• Current Assignee: Mitsubishi Denki Kabushiki Kaisha (Mitsubishi Electric Corporation)
• Original Assignee: Mitsubishi Denki Kabushiki Kaisha (Mitsubishi Electric Corporation)
• Inventors: Suzuki, Tadashi
• Primary Examiner(s): MacDonald, Allen R.
• Assistant Examiner(s): EDOUARD, PATRICK NESTOR
• Application Number:
  US08/732,891
• Time in Patent Office: 565 Days
• Field of Search: 704/233, 704/240, 704/227, 704/228, 704/226, 381/94
• US Class Current: 704/233
• CPC Class Codes: G10L 15/20 Speech recognition techniques specially adapted for robustness in adverse environments, e.g. in noise, of stress induced speech G10L21/02 takes precedence

Abstract

An estimated-SN(Signal Noise)-ratio is calculated for a time-series feature vector of noise-superimposed speech by using a noise-free speech model and a noise model. A noise-superimposed model is generated based on the estimated-SN-ratio. A likelihood between the time-series feature vector of noise-superimposed speech and the noise-superimposed model is calculated to obtain likelihood information. A noise spectrum included in the noise-superimposed speech is estimated from the likelihood information.

Claims

1. A speech recognition apparatus for recognizing an input speech under noisy circumstances comprising:

a noise model memory for storing a noise model;

a speech model memory for storing a noise-free speech model;

a reference model memory for storing a plurality of speech models for collation;

an acoustic analyzer for receiving the input speech, acoustically analyzing a noise-superimposed speech signal of the input speech, and outputting a time-series feature vector of noise-superimposed speech;

a superimposed-noise estimating unit for estimating a superimposed noise based on the time-series feature vector of noise-superimposed speech by using the noise model stored in the noise model memory and the noise-free speech model stored in the speech model memory, and outputting an estimated superimposed-noise spectrum;

a spectrum calculator for receiving the input speech, analyzing a spectrum of the noise-superimposed speech signal of the input speech, and outputting a time-series noise-superimposed speech spectrum;

a noise spectrum eliminator for eliminating a spectrum component of a noise speech in the noise-superimposed speech signal for the time-series noise-superimposed speech spectrum output from the spectrum calculator by using the estimated superimposed-noise spectrum output from the superimposed-noise estimating unit, and outputting a time-series noise-eliminated speech spectrum;

a feature vector calculator for calculating a first feature vector from the time-series noise-eliminated speech spectrum and outputting a time-series feature vector of noise-eliminated speech; and

a collating unit for collating the time-series feature vector of noise-eliminated speech with the plurality of speech models for collation stored in the reference model memory, selecting a speech model out of the plurality of speech models for collation, whose likelihood is highest, and outputting the speech model as a recognition result.

2. The noise recognition apparatus of claim 1, wherein the superimposed-noise estimating unit includes:

an estimated-SN(Signal Noise)-ratio calculator for inputting the time-series feature vector of noise-superimposed speech output from the acoustic analyzer, calculating an estimated-SN-ratio for each noise-superimposed speech feature vector by using the noise model stored in the noise model memory and the noise-free speech model stored in the speech model memory, and outputting the estimated-SN-ratio;

a noise-superimposed speech model generator for synthesizing the noise model stored in the noise model memory with the noise-free speech model stored in the speech model memory based on the estimated-SN-ratio, generating a noise-superimposed speech model and outputting the noise-superimposed speech model;

a likelihood calculator for collating the noise-superimposed speech model with the noise-superimposed speech feature vector, which was used for calculating the estimated-SN-ratio in the estimated-SN-ratio calculator, calculating a likelihood between the noise-superimposed speech model and the noise-superimposed speech feature vector, and outputting the likelihood as first collation data;

a weight data calculator for calculating first weight data for a combination of the each noise-superimposed speech feature vector, the noise model and the noise-free speech model by using the first collation data, and outputting the first weight data; and

a noise spectrum generator for generating an estimated superimposed-noise spectrum by using the first weight data, the estimated-SN-ratio output from the estimated-SN-ratio calculator, the time-series feature vector of noise-superimposed speech output from the acoustic analyzer, and the noise model stored in the noise model memory, and for outputting the estimated superimposed-noise spectrum.

3. The noise recognition apparatus of claim 2, wherein the noise model memory stores at least one noise spectrum and a second feature vector corresponding to each of the noise spectrum, as the noise model.

4. The noise recognition apparatus of claim 2, wherein the speech model memory stores at least one noise-free speech feature vector as the noise-free speech model.

5. The noise recognition apparatus of claim 2, wherein the speech model memory stores a model in which syllable models are mutually connected, as the noise-free speech model.

6. The noise recognition apparatus of claim 2, wherein the likelihood calculator calculates second collation data per combination of the noise model and the speech model for collation and outputs the second collation data, and the weight data calculator selects a highest likelihood out of the second collation data output from the likelihood calculator to calculate second weight data.

7. The noise recognition apparatus of claim 2, wherein the likelihood calculator calculates second collation data per combination of the noise model and the speech model for collation and outputs the second collation data, and the weight data calculator selects a plurality of likelihoods, in order of highness of the likelihood, out of the second collation data output from the likelihood calculator to calculate third weight data.

8. The noise recognition apparatus circumstances of claim 7, wherein the likelihood calculator calculates the second collation data per combination of the noise model and the speech model for collation and outputs the second collation data, and the weight data calculator selects the plurality of likelihoods, in order of highness of the likelihood, out of the second collation data, performs a weighted addition of likelihood obtained from the noise model used for calculating the second collation data, and calculates fourth weight data by obtaining one of the noise model whose result of the weighted addition is maximum.

9. A speech recognition method for recognizing an input speech under noisy circumstances, including a noise model memory for storing a noise model, a speech model memory for storing a noise-free speech model, and a reference model memory for storing a plurality of speech models for collation, the method comprising the steps of:

analyzing a noise-superimposed speech signal of the input speech acoustically to output a time-series feature vector of noise-superimposed speech;

estimating a superimposed-noise for the time-series feature vector of noise-superimposed speech by using the noise model stored in the noise model memory and the noise-free speech model stored in the speech model memory to output an estimated superimposed-noise spectrum;

calculating a spectrum of the noise-superimposed speech signal in the input speech by performing a spectrum-analysis to output a time-series noise-superimposed speech spectrum;

eliminating a spectrum component of a noise speech in the noise-superimposed speech signal for the time-series noise-superimposed speech spectrum output from the step of calculating the spectrum by using the estimated superimposed-noise spectrum output from the step of estimating the superimposed-noise to output a time-series noise-eliminated speech spectrum;

calculating a first feature vector from the time-series noise-eliminated speech spectrum to output a time-series feature vector of noise-eliminated speech; and

collating the time-series feature vector of noise-eliminated speech with the plurality of speech models for collation stored in the reference model memory to select and output a speech model out of the plurality of speech models for collation, whose likelihood is highest, as a recognition result.

10. The noise recognition method of claim 9, wherein the step of estimating the superimposed-noise further includes steps of:

calculating an estimated-SN-ratio for each noise-superimposed speech feature vector of the time-series feature vector of noise-superimposed speech output from the step of analyzing by using the noise model stored in the noise model memory and the noise-free speech model stored in the speech model memory to output the estimated-SN-ratio;

generating a noise-superimposed speech model by synthesizing the noise model stored in the noise model memory with the noise-free speech model stored in the speech model memory based on the estimated-SN-ratio to output the noise-superimposed speech model;

calculating a likelihood by collating the noise-superimposed speech model with the noise-superimposed speech feature vector, which was used for calculating the estimated-SN-ratio in the step of calculating the estimated-SN-ratio, and by calculating the likelihood between the noise-superimposed speech model and the noise-superimposed speech feature vector to output the likelihood as first collation data;

calculating first weight data for a combination of the each noise-superimposed speech feature vector, the noise model and the noise-free speech model by using the first collation data to output the first weight data; and

generating an estimated superimposed-noise spectrum by using the first weight data, the estimated-SN-ratio output from the step of calculating the estimated-SN-ratio, the time-series feature vector of noise-superimposed speech output from the step of analyzing, and the noise model stored in the noise model memory to output the estimated superimposed-noise spectrum.

11. The noise recognition method of claim 10, wherein the step of calculating the likelihood calculates second collation data per combination of the noise model and the speech model for collation and outputs the second collation data, and the step of calculating the weight data selects a highest likelihood out of the second collation data to calculate second weight data.

12. The noise recognition method of claim 10, wherein the step of calculating the likelihood calculates second collation data per combination of the noise model and the speech model and outputs the second collation data, and the step of calculating the weight data selects a plurality of likelihoods, in order of highness of the likelihood, out of the second collation data to calculate third weight data.

13. The noise recognition method of claim 12, wherein the step of calculating the likelihood calculates the second collation data per combination of the noise model and the speech model for collation and outputs the second collation data, and the step of calculating the weight data selects the plurality of likelihoods, in order of highness of the likelihood, out of the second collation data, performs a weighted addition of likelihood obtained from the noise model used for calculating the second collation data, and calculates fourth weight data by obtaining one of the noise model whose result of the weighted addition is maximum.