Noise reduction system using neural network

• US 5,185,848 A
• Filed: 12/12/1989
• Issued: 02/09/1993
• Est. Priority Date: 12/14/1988
• Status: Expired due to Fees

• Current Assignee: Hitachi, Ltd.
• Original Assignee: Hitachi, Ltd.
• Inventors: Aritsuka, Toshiyuki, Amano, Akio, Hataoka, Nobuo, Ichikawa, Akira
• Primary Examiner(s): Fleming, Michael R.
• Assistant Examiner(s): DAVIS, GEORGE B
• Application Number:
  US07/448,949
• Time in Patent Office: 1,155 Days
• Field of Search: 364/513, 381/43, 395/22, 395/2
• US Class Current: 704/202
• CPC Class Codes: G10L 15/16 using artificial neural networks
• CPC Class Codes: G10L 21/0208 Noise filtering
• CPC Class Codes: G10L 21/0364 for improving intelligibility

Abstract

A noise reduction system used for transmission and/or recognition of speech includes a speech analyzer for analyzing a noisy speech input signal thereby converting the speech signal into feature vectors such as autocorrelation coefficients, and a neural network for receiving the feature vectors of the noisy speech signal as its input. The neural network extracts from a codebook an index of prototype vectors corresponding to a noise-free equivalent to the noisy speech input signal. Feature vectors of speech are read out from the codebook on the basis of the index delivered as an output from the neural network, thereby causing the speech input to be reproduced on the basis of the feature vectors of speech read out from the codebook.

Claims

1. A noise reduction system for transmitting a noise-suppressed speech signal, comprising:

speech analysis means for a analyzing a noisy speech input signal thereby converting the speech signal into feature vectors;

maximum detector means for detecting a maximum value from among the elements of said feature vectors obtained by said speech analysis means;

normalization means for normalizing values of the elements of said feature vectors on the basis of the maximum value detected by said maximum detector means;

a neural network for receiving said normalized feature vectors and extracting an index of a representative vector corresponding to a noise-free speech signal equivalent to said noisy speech input signal from a codebook generated by previously clustering a set of feature vectors of the noise-free speech signal and storing representative vectors of said feature vectors of the noise-free speech signal together with corresponding indices of said representative vectors; and

transmitter means for transmitting both the maximum value detected by said maximum detector means from among the elements of said feature vectors and the index extracted from said codebook by said neural network.

2. A noise reduction system according to claim 1, wherein said maximum detector means detects the maximum value from among the elements of the feature vectors which are adversely affected less by noise among those obtained by said speech analysis means.

3. A noise reduction system according to claim 1, wherein said codebook is generated by previously preparing noise and a noise-free speech signal, converting the noise-free speech signal into feature vectors, clustering said feature vectors and selecting representative vectors representing the clusters of said feature vectors, and said neural network is trained using as teacher data the index of the prototype vectors representing a most appropriate cluster corresponding to said feature vectors and as learning data the feature vectors obtained by converting a noise-superposed speech signal consisting of the noise and the noise-free speech signal.

4. A noise reduction system according to claim 1, wherein said speech analysis means converts said noisy speech signal into autocorrelation coefficients.

5. A noise reduction system according to claim 4, wherein said maximum detector means detects the maximum value from among said autocorrelation coefficients except that of zeroth order which is most adversely affected by noise.

6. A noise reduction system according to claim 1 wherein said neural network has a hierarchial structure.

7. A noise reduction system according to claim 1, wherein said neural network has a structure of Hopfield type.

8. A noise reduction system according to claim 1, wherein said neutral network has a structure of Boltzmann'"'"'s machine type.

9. A noise reduction system according to claim 1, wherein said neural network has a structure of multi-stage type in which a plurality of neural networks are arranged in series and in parallel.

10. A speech coding and decoding system comprising:

speech analysis means for analyzing a noisy speech input signal thereby converting the speech signal into feature vectors;

maximum detector means for detecting a maximum value from among the elements of said feature vectors obtained by said speech analysis means;

normalization means for normalizing the values of the elements of said feature vectors on the basis of the maximum value detected by said maximum detector means;

a neural network for receiving said normalized feature vectors and extracting an index of a representative vector corresponding to a noise-free speech signal equivalent to said noisy speech signal from a codebook generated by previously clustering a set of feature vectors of the noise-free speech signal and storing representative vectors of said features vectors of the noise-free speech input signal together with corresponding indices of said representative vectors;

transmitter means for transmitting both the maximum value detected by said maximum detector means from among the elements of said feature vectors and the index extracted from said codebook by said neural network;

receiver means for receiving and decoding information transmitted from said transmitter means;

vector selector means for receiving the transmitted index to select the representative vector corresponding to said index from said codebook and generating said selected representative vector as its output, said codebook storing representative vectors suitable for speech synthesis together with corresponding indices; and

speech synthesizer means for synthesizing speech on the basis of said maximum value and said representative vector generated by said vector selector means as being suitable for the speech synthesis.

11. A speech coding and decoding system according to claim 10, wherein said speech analysis means converts said noisy speech signal into autocorrelation coefficients.

12. A speech coding and decoding system according to claim 10, wherein said representative vectors suitable for speech synthesis are partial autocorrelation coefficients.

13. A noise reduction system according to claim 10, wherein said neural network has a hierarchial structure.

14. A noise reduction system according to claim 10, wherein said neural network has a structure of Hopfield type.

15. A noise reduction system according to claim 10, wherein said neutral network has a structure of Boltzmann'"'"'s machine type.

16. A noise reduction system according to claim 10, wherein said neural network has a structure of multi-stage type in which a plurality of neural networks are arranged in series and in parallel.