Adaptive acoustic echo delay estimation

1. A method for acoustic echo cancellation, comprising:

receiving, via a communication network, a first acoustic signal from a far-end environment;

receiving, by a microphone, a second acoustic signal from a near-end environment, the second acoustic signal including a delayed version of the first acoustic signal;

modeling, by a processor, a relationship between the first acoustic signal and the second acoustic signal using an adaptive filter, the adaptive filter using sampling points of the first acoustic signal and the second acoustic signal along a timeline as inputs;

identifying, by the processor, a sampling point among the sampling points, wherein weight values of the adaptive filter associated with the identified sampling point experience an increase above a threshold value and wherein the identified sampling point along the timeline represents an estimated delay between the first acoustic signal and the second acoustic signal; and

removing, by the processor, the delayed version of the first acoustic signal from the second acoustic signal based on the estimated delay.

2. The method of claim 1 , wherein the identifying the sampling point comprises:

identifying, by the processor, a sampling point i among the sampling points, wherein

i

:=

w

⁡

(

i

)

max

⁡

(

w

⁡

(

1

⁢

:

⁢

i

-

1

)

)

>

threshold

,

and wherein w(1: i−1) refers to weight values associated with each individual sample point prior to the identified sampling point i along the time line, w(i) refers to weight values associated with the identified sampling point i along the time line, and threshold refers to the threshold value.

3. The method of claim 1 , wherein the threshold value is within a range from 1.2 to 5.

4. The method of claim 1 , wherein the second acoustic signal further includes an audio generated by an acoustic source of the near-end environment; and

wherein the method further comprises: transmitting, via the communication network, to the far-end environment the second acoustic signal with the delayed version of the first acoustic signal removed.

5. The method of claim 1 , wherein the modeling using the adaptive filter comprises:

for a sampling point, estimating a difference between the second acoustic signal and the first acoustic signal being modified by the adaptive filter;

updating filter weights of the adaptive filter based on the estimated difference; and

iterating the estimating and the updating for additional sampling points until a convergence condition is met.

6. The method of claim 5 , further comprising:

normalizing the difference between the second acoustic signal and the first acoustic signal being modified by the adaptive filter.

7. The method of claim 5 , wherein the updating of the filter weights comprises:

updating filter weights of the adaptive filter based on the estimated difference such that a difference between the second acoustic signal and the first acoustic signal being modified by the adaptive filter for a next sample point is reduced.

8. The method of claim 5 , wherein the convergence condition is that the difference between the second acoustic signal and the first acoustic signal being modified by the adaptive filter is less than 1% of the first acoustic signal being modified by the adaptive filter.

9. The method of claim 5 , wherein the updating of the filter weights comprises: updating filter weights of the adaptive filter based on the estimated difference and a momentum term, and wherein the momentum term depends on an update to the filter weights of the adaptive filter during a previous iteration.

10. The method of claim 5 , wherein the updating of the filter weights comprises: updating filter weights of the adaptive filter based on the estimated difference and a learning factor that varies with time, and wherein the learning factor that varies with time is a time- decaying learning factor that decays as a number of iterations of the estimating and the updating increases.

11. The method of claim 1 , further comprising:

outputting, by a speaker, the first acoustic signal received from the communication network to the near-end environment;

wherein the first acoustic signal received from the communication network is reflected by an object in the near-end environment, after being output by the speaker, before being received by the microphone as the delayed version of the first acoustic signal included in the second acoustic signal.

12. The method of claim 1 , wherein the second acoustic signal further includes a noise from the near-end environment.

13. A method for acoustic echo cancellation, comprising:

receiving, via a communication network, a first acoustic signal from a far-end environment;

receiving, by a microphone, a second acoustic signal from a near-end environment, the second acoustic signal including a delayed version of the first acoustic signal;

modeling, by a processor, a relationship between the first acoustic signal and the second acoustic signal using an adaptive filter including filter weights, the adaptive filter using sampling points of the first acoustic signal and the second acoustic signal along a timeline to update the filter weights in an iterative process based on a momentum term or a time-decaying learning factor;

identifying, by the processor, a sampling point among the sampling points, wherein weight values of the adaptive filter associated with the identified sampling point experience an increase above a threshold value, and wherein the identified sampling point along the timeline represents an estimated delay between the first acoustic signal and the second acoustic signal; and

removing, by the processor, the delayed version of the first acoustic signal from the second acoustic signal based on the estimated delay.

14. The method of claim 13 , wherein the momentum term or the time-decaying learning factor accelerates a convergence of the iterative process of updating the filter weights of the adaptive filter.

15. The method of claim 13 , wherein the filter weights of the adaptive filter is updated based on the momentum term that depends on an update to the filter weights of the adaptive filter during a previous iteration.

16. The method of claim 13 , wherein the filter weights of the adaptive filter is updated based on the time-decaying learning factor that decays as a number of iterations of the iterative process increases.

17. A system for echo cancellation, comprising:

a networking component configured to receive a first acoustic signal from a far-end environment;

a microphone configured to receive a second acoustic signal from a near-end environment, the second acoustic signal including a delayed version of the first acoustic signal; and

a processor, when in operation, configured to conduct a process including:

modeling a relationship between the first acoustic signal and the second acoustic signal using an adaptive filter, the adaptive filter using sampling points of the first acoustic signal and the second acoustic signal along a timeline to update filter weights of the adaptive filter;

identifying a sampling point among the sampling points, wherein weight values of the adaptive filter associated with the identified sampling point experience an increase above a threshold value, and wherein the identified sampling point along the timeline represents an estimated delay between the first acoustic signal and the second acoustic signal; and

removing the delayed version of the first acoustic signal from the second acoustic signal based on the estimated delay.

18. The system of claim 17 , further comprising:

a speaker configured to output the first acoustic signal received from the communication network to the near-end environment.

19. The system of claim 17 , wherein the networking component is further configured to transmit to the far-end environment the second acoustic signal with the delayed version of the first acoustic signal removed.

20. The system of claim 17 , wherein the filter weights are updated in an iterative process based on a momentum term or a time-decaying learning factor, and wherein the momentum term or the time-decaying learning factor accelerates a convergence of the iterative process of updating the filter weights of the adaptive filter.