Split-domain speech signal enhancement

Patent metadata

Field	Value
Publication number	US-10741192-B2
Application number	US-201815973214-A
Country	US
Kind code	B2
Filing date	May 7, 2018
Priority date	May 7, 2018
Publication date	Aug 11, 2020
Grant date	Aug 11, 2020

Abstract

A method and an apparatus for estimating speech signal in split-domain is disclosed. The method includes performing LP analysis on a noisy speech signal to generate a first plurality of LPC and a first residual signal. The method also includes estimating speech LPC spectrum to generate cleaned LPC. The method further includes estimating speech residual spectrum to generate cleaned residual signal. The method also includes synthesizing output signals based on the cleaned LPC and the cleaned residual signal.

First claim

What is claimed is:

1. A method for estimating speech signal at an electronic device, the method comprising:
receiving, at a microphone, input signals, wherein the input signals include at least a noise signal component and a speech signal component;
determining, by the electronic device, whether to perform a first filtering operation based on a characteristic of the input signals;
performing, by the electronic device, the first filtering operation on a first portion of the input signals to generate a plurality of first linear predictive filter coefficients (LPC) and a first residual signal;
calculating, by the electronic device, frequency response of the plurality of the first LPC to generate a first magnitude spectrum and a first phase spectrum, wherein the first magnitude spectrum corresponds to magnitude component of the frequency response and the first phase spectrum corresponds to phase component of the frequency response;
converting, by the electronic device, the first residual signal into frequency-domain signal to generate a second magnitude spectrum and a second phase spectrum, wherein the second magnitude spectrum corresponds to magnitude component of the first residual signal in frequency domain and the second phase spectrum corresponds to phase component of the first residual signal in frequency domain;
estimating, by the electronic device, a third magnitude spectrum based on the first magnitude spectrum, wherein the third magnitude spectrum corresponds to the speech signal component;
estimating, by the electronic device, a fourth magnitude spectrum based on the second magnitude spectrum, wherein the fourth magnitude spectrum corresponds to the speech signal component; and
synthesizing output signals, by the electronic device, based on the third magnitude spectrum and the fourth magnitude spectrum.

2. The method of claim 1 , wherein synthesizing the output signals comprises:
calculating, by the electronic device, a plurality of second linear predictive filter coefficients (LPC) based on the third magnitude spectrum;
and
performing, by the electronic device, a second filtering operation based at least in part on the plurality of the second LPC to generate the output signals.

3. The method of claim 2 , wherein synthesizing the output signals comprises converting, by the electronic device, the fourth magnitude spectrum into time-domain signal to generate a second residual signal, wherein the second filtering operation to generate the output signals is based on the second residual signal.

4. The method of claim 1 , wherein estimating the third magnitude spectrum is based on one among a non-negative matrix factorization technique and a neural network based technique.

5. The method of claim 1 , wherein estimating the fourth magnitude spectrum is based on one among a non-negative matrix factorization technique and a neural network based technique.

6. The method of claim 1 , wherein estimating the third magnitude spectrum comprises estimating a plurality of weights based at least on one among a speech dictionary and a noise dictionary trained in linear predictive filter coefficients (LPC) domain.

7. The method of claim 1 , wherein estimating the fourth magnitude spectrum comprises estimating a plurality of weights based at least on one among a speech dictionary and a noise dictionary trained in residual signal domain.

8. The method of claim 7 , wherein at least one weight of the plurality of weights is perceptually weighted or filtered to enhance periodicity.

9. The method of claim 2 , wherein calculating the plurality of the second LPC is further based on the first phase spectrum.

10. The method of claim 3 , wherein converting the fourth magnitude spectrum into time-domain signal is further based on the second phase spectrum.

11. The method of claim 2 , wherein the first filtering operation corresponds to linear predictive analysis filtering and the second filtering operation corresponds to linear predictive synthesis filtering.

12. The method of claim 6 , wherein estimating the third magnitude spectrum comprises:
estimating a first plurality of weight vector based on the speech dictionary; and
estimating a second plurality of weight vector based on the noise dictionary, wherein the third magnitude spectrum is based on the first plurality of weight vector.

13. The method of claim 6 , wherein estimating the fourth magnitude spectrum comprises:
estimating a third plurality of weight vector based on the speech dictionary; and
estimating a fourth plurality of weight vector based on the noise dictionary, wherein the fourth magnitude spectrum is based on the third plurality of weight vector.

14. An apparatus for estimating speech signal, comprising:
a microphone configured to receive input signals, wherein the input signals include at least a noise signal component and a speech signal component;
a memory configured to store the input signals; and
a processor coupled to the memory, the processor configured to:
perform a first filtering operation on a first portion of the input signals to generate a plurality of first linear predictive filter coefficients (LPC) and a first residual signal;
calculate frequency response of the plurality of the first LPC to generate a first magnitude spectrum and a first phase spectrum, wherein the first magnitude spectrum corresponds to magnitude component of the frequency response and the first phase spectrum corresponds to phase component of the frequency response;
convert the first residual signal into frequency-domain signal to generate a second magnitude spectrum and a second phase spectrum, wherein the second magnitude spectrum corresponds to magnitude component of the first residual signal in frequency domain and the second phase spectrum corresponds to phase component of the first residual signal in frequency domain;
estimate a third magnitude spectrum based on the first magnitude spectrum, wherein the third magnitude spectrum corresponds to the speech signal component;
estimate a fourth magnitude spectrum based on the second magnitude spectrum, wherein the fourth magnitude spectrum corresponds to the speech signal component;
convert, based on the second phase spectrum, the fourth magnitude spectrum into time-domain signal to generate a second residual signal; and
synthesize output signals based on the third magnitude spectrum and the second residual signal.

15. The apparatus of claim 14 , wherein the processor is further configured to determine whether to perform the first filtering operation based on a characteristic of the input signals.

16. The apparatus of claim 14 , wherein the processor is configured to synthesize the output signals based on a plurality of second linear predictive filter coefficients (LPC) that is based on the third magnitude spectrum.

17. The apparatus of claim 14 , wherein the processor is configured to estimate the third magnitude spectrum based on one among a non-negative matrix factorization technique and a neural network based technique.

18. The apparatus of claim 14 , wherein the processor is configured to estimate the fourth magnitude spectrum based on one among a non-negative matrix factorization technique and a neural network based technique.

19. The apparatus of claim 14 , wherein the processor is further configured to estimate a plurality of weights based at least on one among a speech dictionary and a noise dictionary trained in linear predictive filter coefficients (LPC) domain.

20. The apparatus of claim 14 , wher

Assignees

• Qualcomm Inc

Inventors

• Rajendran Vivek
• Dewasurendra Duminda
• Sinder Daniel Jared

Classifications

• G10L25/90

  Pitch determination of speech signals · CPC title

• G10L25/12

  the extracted parameters being prediction coefficients · CPC title

• G10L21/0208 Primary

  Noise filtering · CPC title

• G10L19/08

  Determination or coding of the excitation function; Determination or coding of the long-term prediction parameters · CPC title

• G10L25/30

  using neural networks · CPC title