Heterogeneous computing for hybrid acoustic echo cancellation

1. A device operative to perform acoustic echo cancellation, comprising:

a speaker to output a far-end signal at the device;

a microphone to receive at least a near-end signal and the far-end signal from the speaker to produce a microphone output;

an artificial intelligence (AI) accelerator operative to perform neural network (NN) operations according to a first NN model and a second NN model to output an echo-suppressed signal; and

a digital signal processing (DSP) unit operative to:

perform adaptive filtering to remove at least a portion of the far-end signal from the microphone output to generate a filtered near-end signal; and

perform Fast Fourier Transform (FFT) and inverse FFT (IFFT) to generate input to the first NN model and the second NN model, respectively.

2. The device of claim 1 , wherein the DSP unit in performing the FFT is further operative to:

compute spectra of the far-end signal and the filtered near-end signal to generate input to the first NN model.

3. The device of claim 1 , wherein the DSP unit in performing the IFFT is further operative to:

perform the IFFT on an output of the first NN model to generate a time-domain signal as input to the second NN model.

4. The device of claim 1 , further comprising:

a pipeline including:

two or more DSP stages to perform the adaptive filtering, the FFT, and the IFFT; and

two or more NN stages to perform the NN operations according to the first NN model and the second NN model.

5. The device of claim 1 , further comprising:

a 5-stage pipeline including:

a first DSP circuit to perform the adaptive filtering;

a second DSP circuit to perform the FFT;

a first NN circuit to operate according to the first NN model;

a third DSP circuit to perform the IFFT; and

a second NN circuit to operate according to the second NN model.

6. The device of claim 1 , wherein the far-end signal is input to the FFT and the second NN model.

7. The device of claim 1 , wherein the AI accelerator is operative to perform operations according to the first NN model on a normalized spectrum of the far-end signal concatenated with a normalized spectrum of the filtered near-end signal.

8. The device of claim 1 , wherein the AI accelerator is operative to perform operations according to the second NN model on a segmented far-end signal and an output of the IFFT.

9. The device of claim 1 , wherein the FFT is a short-time Fourier transformation (STFT).

10. The device of claim 1 , wherein both the first NN model and the second NN model each includes two layers of long short term memory (LSTM), a fully-connected layer, and sigmoid activation.

11. A device including a multi-stage pipeline comprising:

two or more digital signal processing (DSP) stages to perform adaptive filtering, Fast Fourier Transform (FFT), and inverse FFT (IFFT); and

two or more neural network stages to perform neural network operations according to a first NN model and a second NN model,

wherein the adaptive filtering removes at least a portion of a far-end signal from a microphone output to generate a filtered near-end signal as input to the FFT, and

wherein the second NN model receives an output of the IFFT and generates an echo-suppressed signal.

12. The device of claim 11 , wherein the multi-stage pipeline further comprises:

a first DSP circuit to perform the adaptive filtering;

a second DSP circuit to perform the FFT to compute spectra of the far-end signal and the filtered near-end signal to generate input to the first NN model;

a first NN circuit to operate on the spectra according to the first NN model;

a third DSP circuit to perform the IFFT on an output of the first NN model to generate a time-domain signal as input to the second NN model; and

a second NN circuit to operate according to the second NN model.

13. The device of claim 11 , wherein one of the neural network stages is operative to perform operations according to the first NN model on a normalized spectrum of the far-end signal concatenated with a normalized spectrum of the filtered near-end signal.

14. The device of claim 11 , wherein one of the neural network stages is operative to perform operations according to the second NN model on a segmented far-end signal and the output of the IFFT.

15. The device of claim 11 , wherein the FFT is a short-time Fourier transformation (STFT).

16. The device of claim 11 , wherein both the first NN model and the second NN model each includes two layers of long short term memory (LSTM), a fully-connected layer, and sigmoid activation.

17. A method for acoustic echo cancellation, comprising:

performing adaptive filtering by a digital signal processing (DSP) unit to remove at least a portion of a far-end signal from a microphone output to generate a filtered near-end signal;

performing, by the DSP unit, Fast Fourier Transform (FFT) to compute spectra of the far-end signal and the filtered near-end signal to generate input to a first neural network (NN) model;

performing first NN operations by an artificial intelligence (AI) accelerator on the spectra according to a first NN model;

performing inverse FFT (IFFT) by the DSP unit on an output of the first NN model to generate a time-domain signal as input to a second NN model; and

performing second NN operations by the AI accelerator according to the second NN model to output an echo-suppressed signal.

18. The method of claim 17 , wherein performing the first NN operations further comprises:

performing the first NN operation according to the first NN model on a normalized spectrum of the far-end signal concatenated with a normalized spectrum of the filtered near-end signal.

19. The method of claim 17 , wherein performing the second NN operations further comprises:

performing the second NN operations according to the second NN model on a segmented far-end signal and an output of the IFFT.

20. The method of claim 17 , wherein the adaptive filtering, the FFT, the first NN model, the IFFT, the second NN model are performed in a multi-stage pipeline.