Step-size control for multi-channel acoustic echo canceller

• US 9,754,605 B1
• Filed: 06/09/2016
• Issued: 09/05/2017
• Est. Priority Date: 06/09/2016
• Status: Active Grant

• Current Assignee: Amazon Technologies, Inc. (Amazon.com, Inc.)
• Original Assignee: Amazon Technologies, Inc. (Amazon.com, Inc.)
• Inventors: Chhetri, Amit Singh
• Primary Examiner(s): Islam, Mohammad
• Application Number:
  US15/177,624
• Time in Patent Office: 453 Days
• CPC Class Codes: G10L 2021/02082 the noise being echo, reverberation of the speech
• CPC Class Codes: G10L 21/0264 characterised by the type of parameter measurement, e.g. correlation techniques, zero crossing techniques or predictive techniques
• CPC Class Codes: G10L 25/06 the extracted parameters being correlation coefficients
• CPC Class Codes: G10L 25/21 the extracted parameters being power information
• CPC Class Codes: H04R 2499/11 Transducers incorporated or for use in hand-held devices, e.g. mobile phones, PDA's, camera's
• CPC Class Codes: H04R 3/005 for combining the signals of two or more microphones specially adapted for hearing aids H04R25/407
• CPC Class Codes: H04R 3/02 for preventing acoustic reaction , i.e. acoustic oscillatory feedback specially adapted for hearing aids H04R25/453
• CPC Class Codes: H04S 7/305 Electronic adaptation of stereophonic audio signals to reverberation of the listening space H04S7/301 takes precedence

Abstract

A multi-channel acoustic echo cancellation (AEC) system that includes a step-size controller that dynamically determines a step-size value for each channel and each tone index on a frame-by-frame basis. The system determines the step-size value based on a normalized squared cross-correlation (NSCC) between an estimated echo signal and an error signal, allowing the AEC system to converge quickly when an acoustic room response changes while providing stable steady-state error by avoiding misadjustments due to noise sensitivity and/or near-end speech. The step-size value can be determined using fractional weighting that takes into account a signal strength of each channel.

Claims

1. A computer-implemented method implemented on a voice-controllable device, the method determining a step-size value of a first adaptive filter of the device, the method comprising:

receiving a first reference audio signal that is sent from the device to a first loudspeaker for audio playback;

receiving, from a microphone of the device, a first microphone audio signal representing audible sound output by the first loudspeaker;

determining, using the first reference audio signal and the first adaptive filter that is configured to adjust according to an optimization algorithm, a first echo audio signal that is an estimated representation of a portion of the first microphone audio signal;

determining a plurality of echo audio signals;

determining a combined echo audio signal by summing the plurality of echo audio signals and the first echo audio signal;

determining an error signal by subtracting the combined echo audio signal from the first microphone audio signal;

determining a first normalized squared cross-correlation (NSCC) value between the error signal and the first echo audio signal;

determining a first scale factor using the first NSCC value, the first scale factor becoming larger as the first NSCC value approaches a value of one;

determining a first weight corresponding to a magnitude of the first reference audio signal;

determining the step-size value by multiplying the first scale factor, the first weight and a nominal step-size value, the step-size value corresponding to the first reference audio signal; and

providing the step-size value to the first adaptive filter.

2. The computer-implemented method of claim 1, wherein determining the first scale factor further comprises:

determining a first power value corresponding to the first echo audio signal;

determining second power value corresponding to the error signal;

determining a first product by multiplying one plus the first NSCC value by the first power value;

determining a second product by multiplying one minus the first NSCC value by the second power value;

determining a first sum by adding the first power value to the second product; and

determining the first scale factor by dividing the first product by the first sum.

3. The computer-implemented method of claim 1, wherein determining the first NSCC value further comprises:

determining a first smoothing value between zero and one, the first smoothing value indicating a weight associated with a first cross-correlation value at a first time;

determining a second smoothing value by subtracting the first smoothing value from one;

determining the first cross-correlation value between the error signal and the first echo audio signal at the first time;

generating a first product by multiplying the first smoothing value and the first cross-correlation value;

generating a second product by multiplying the second smoothing value, the first echo audio signal and the error signal;

determining a second cross-correlation value between the error signal and the first echo audio signal at a second time after the first time by summing the first product and the second product; and

determining the first normalized cross-correlation value by normalizing the second cross-correlation value.

4. The computer-implemented method of claim 1, wherein determining the first weight further comprises:

determining a first portion of the first reference audio signal that corresponds to a first duration of time and a first frequency range;

determining a first portion of the second reference audio signal that corresponds to the first duration of time and the first frequency range;

determining a first power value corresponding to a magnitude of the first portion of the first reference audio signal;

determining a second power value corresponding to a magnitude of the first portion of the second reference audio signal;

determining that the second power value is greater than the first power value; and

determining the first weight by dividing the first power value by the second power value.

5. A computer-implemented method, comprising:

receiving a first reference signal corresponding to a first audio channel;

receiving a second reference signal corresponding to a second audio channel;

receiving a first audio input signal;

determining, using a first adaptive filter and the first reference signal, a first echo signal that models a first portion of the first audio input signal;

determining, using a second adaptive filter and the second reference signal, a second echo signal that models a second portion of the first audio input signal;

combining the first echo signal and the second echo signal to generate a combined echo signal;

determining an error signal by subtracting the combined echo signal from the first audio input signal;

determining a first normalized squared cross-correlation (NSCC) value associated with the error signal and the first echo signal;

determining a first scale factor based on the first NSCC value; and

determining a first step-size value based on the first scale factor and a nominal step-size value, the first step-size value corresponding to the first reference signal.

6. The computer-implemented method of claim 5, wherein the first step-size value corresponds to the first reference signal, a first duration of time, and a first frequency range, and the method further comprises:

determining a second step-size value, the second step-size value corresponding to the first reference signal, the first duration of time and a second frequency range;

determining a third step-size value, the third step-size value corresponding to the second reference signal, the first duration of time and the first frequency range;

sending the first step-size value to the first adaptive filter;

sending the second step-size value to the first adaptive filter;

sending the third step-size value to the second adaptive filter; and

performing acoustic echo cancellation using the first adaptive filter and the second adaptive filter.

7. The computer-implemented method of claim 5, wherein determining the first scale factor further comprises:

determining a first power value corresponding to the first echo signal;

determining second power value corresponding to the error signal;

determining a first product by multiplying the first NSCC value by a first constant;

determining a second product by multiplying one plus the first product by the first power value;

determining a third product by multiplying one minus the first NSCC value by the second power value;

determining a first sum by adding the first power value to the third product; and

determining the first scale factor by dividing the second product by the first sum.

8. The computer-implemented method of claim 5, wherein determining the first NSCC value further comprises:

determining a first smoothing value between zero and one, the first smoothing value indicating a weight associated with a first cross-correlation value that corresponds to a first time;

determining a second smoothing value by subtracting the first smoothing value from one;

determining the first cross-correlation value between the error signal and the first echo signal at the first time, the first cross-correlation value corresponding to a second frame preceding the first frame;

generating a first product by multiplying the first smoothing value and the first cross-correlation value;

generating a second product by multiplying the second smoothing value, the first echo signal and the error signal;

determining a second cross-correlation value between the error signal and the first echo signal at a second time after the first time by summing the first product and the second product; and

determining the first NSCC value by normalizing the second cross-correlation value.

9. The computer-implemented method of claim 8, wherein determining the first NSCC value further comprises:

determining a first power value corresponding to the first echo signal;

determining a second power value corresponding to the error signal;

determining a third product by multiplying the first power value by the second power value;

determining a first denominator by taking a square root of the third product;

determining a first value by dividing the second cross-correlation value by the denominator; and

determining the first NSCC value by squaring a magnitude of the first value.

10. The computer-implemented method of claim 5, further comprising:

determining a first weight corresponding to a magnitude of the first reference signal; and

determining the first step-size value based on the first scale factor, the first weight and the nominal step-size.

11. The computer-implemented method of claim 10, wherein determining the first weight further comprises:

determining a first portion of the first reference signal that corresponds to a first duration of time and a first frequency range;

determining a first portion of the second reference signal that corresponds to the first duration of time and the first frequency range;

determining a first power value corresponding to a magnitude of the first portion of the first reference signal;

determining a second power value corresponding to a magnitude of the first portion of the second reference signal;

determining that the second power value is greater than the first power value; and

determining the first weight by dividing the first power value by the second power value.

12. The computer-implemented method of claim 5, wherein determining the first echo signal further comprises:

estimating a first transfer function corresponding to an impulse response;

determining a weight vector based on the first transfer function, the weight vector corresponding to adaptive filter coefficients; and

determining the first echo signal by convolving the first reference signal with the weight vector.

13. A first device, comprising:

at least one processor;

a wireless transceiver; and

a memory device including first instructions operable to be executed by the at least one processor to configure the first device to;

receive a first reference signal corresponding to a first audio channel;

receive a second reference signal corresponding to a second audio channel;

receive a first input signal;

determine, using a first adaptive filter and the first reference signal, a first echo signal that models a first portion of the first audio input signal;

determine, using a second adaptive filter and the second reference signal, a second echo signal that models a second portion of the first audio input signal;

combining the first echo signal and the second echo signal to generate a combined echo signal;

determine an error signal by subtracting the combined echo signal from the first audio input signal;

determine a first normalized squared cross-correlation (NSCC) value associated with the error signal and the first echo signal;

determine a first scale factor based on the first NSCC value; and

determine a first step-size value based on the first scale factor and a nominal step-size value, the first step-size value corresponding to the first reference signal.

14. The first device of claim 13, wherein the first step-size value corresponds to the first reference signal, a first duration of time and a first frequency range, and the second instructions further configure the first device to:

determine a second step-size value, the second step-size value corresponding to the first reference signal, the first duration of time and a second frequency range;

determine a third step-size value, the third step-size value corresponding to the second reference signal, the first duration of time and the first frequency range;

send the first step-size value to the first adaptive filter;

send the second step-size value to the first adaptive filter;

send the third step-size value to the second adaptive filter; and

perform acoustic echo cancellation using the first adaptive filter and the second adaptive filter.

15. The first device of claim 13, wherein the second instructions further configure the first device to:

determine a first power value corresponding to the first echo signal;

determine second power value corresponding to the error signal;

determine a first product by multiplying the first NSCC value by a first constant;

determine a second product by multiplying one plus the first product by the first power value;

determine a third product by multiplying one minus the first NSCC value by the second power value;

determine a first sum by adding the first power value to the third product; and

determine the first scale factor by dividing the second product by the first sum.

16. The first device of claim 13, wherein the second instructions further configure the first device to:

determine a first smoothing value between zero and one, the first smoothing value indicating a weight associated with a first cross-correlation value that corresponds to a first time;

determine a second smoothing value by subtracting the first smoothing value from one;

determine the first cross-correlation value between the error signal and the first echo signal at the first time, the first cross-correlation value corresponding to a second frame preceding the first frame;

generate a first product by multiplying the first smoothing value and the first cross-correlation value;

generate a second product by multiplying the second smoothing value, the first echo signal and the error signal;

determine a second cross-correlation value between the error signal and the first echo signal at a second time after the first time by summing the first product and the second product; and

determine the first NSCC value by normalizing the second cross-correlation value.

17. The first device of claim 16, wherein the second instructions further configure the first device to:

determine a first power value corresponding to the first echo signal;

determine a second power value corresponding to the error signal;

determine a third product by multiplying the first power value by the second power value;

determine a first denominator by taking a square root of the third product;

determine a first value by dividing the second cross-correlation value by the denominator; and

determine the first NSCC value by squaring a magnitude of the first value.

18. The first device of claim 13, wherein the second instructions further configure the first device to:

determine a first weight corresponding to a magnitude of the first reference signal; and

determine the first step-size value based on the first scale factor, the first weight and the nominal step-size.

19. The first device of claim 18, wherein the second instructions further configure the first device to:

determine a first portion of the first reference signal that corresponds to a first duration of time and a first frequency range;

determine a first portion of the second reference signal that corresponds to the first duration of time and the first frequency range;

determine a first power value corresponding to a magnitude of the first portion of the first reference signal;

determine a second power value corresponding to a magnitude of the first portion of the second reference signal;

determine that the second power value is greater than the first power value; and

determine the first weight by dividing the first power value by the second power value.

20. The first device of claim 13, wherein the second instructions further configure the first device to:

estimate a first transfer function corresponding to an impulse response;

determine a weight vector based on the first transfer function, the weight vector corresponding to adaptive filter coefficients; and

determine the first echo signal by convolving the first reference signal with the weight vector.