System and method for acoustic echo cancelation using deep multitask recurrent neural networks

1. A system comprising:

a processor; and

memory storing instructions that, when executed by the processor, cause the processor to:

receive a far-end signal and a near-end signal comprising at least an echo signal corresponding to the far-end signal;

generate, by a contextual-attention neural network, a plurality of estimated near-end features based at least in part on the received near-end signal and the received far-end signal, without using an estimated echo signal; and

generate, based on the plurality of estimated near-end features, an estimated near-end signal, the estimated near-end signal corresponding to the near-end signal without the echo signal.

2. The system of claim 1 , wherein the memory further stores instructions that, when executed by the processor, cause the processor to transmit the estimated near-end signal to a far-end device.

3. The system of claim 1 , wherein the far-end signal is received from a speaker device and the near-end signal is received from a microphone device that generates signals from the speaker device.

4. The system of claim 1 , wherein the memory further stores instructions that, when executed by the processor, cause the processor to:

generate, based on the far-end signal, a plurality of far-end features; and

generate, based on the near-end signal, a plurality of near-end features,

wherein the plurality of estimated near-end features are further generated by inputting one or more near-end features of the plurality of near-end features and one or more far-end features of the plurality of far-end features to the contextual-attention neural network.

5. The system of claim 4 , wherein:

the contextual-attention neural network further comprises an encoder configured to encode one or more near-end features and one or more far-end features into encoded features; and

the plurality of near-end features are generated from the encoded features.

6. The system of claim 5 , wherein:

the contextual-attention neural network further comprises a decoder configured to decode one or more encoded features.

7. The system of claim 6 , wherein the encoder and decoder comprise a plurality of gated recurrent units for encoding and decoding signal features.

8. The system of claim 6 , wherein:

the contextual-attention neural network further comprises a contextual-attention neural network module configured to generate one or more contextual-attention features from the encoded features; and

the decoder is configured to decode the one or more contextual-attention features to generate the plurality of near-end features.

9. The system of claim 8 , wherein:

the encoded features are supplied to a multi-head self-attention layer of the contextual-attention neural network module; and

the contextual-attention neural network module further comprises a multi-head attention layer configured to receive an output of the multi-head self-attention layer.

10. The system of claim 9 , wherein the contextual-attention neural network module further comprises a first layer normalization layer between the multi-head self-attention layer and the multi-head attention layer.

11. The system of claim 10 , wherein the contextual-attention neural network module further comprises a second layer normalization layer configured to compute the contextual-attention features in accordance with outputs from other layers of the contextual-attention neural network module.

12. The system of claim 1 , wherein the memory further stores instructions that, when executed by the processor, cause the processor to:

generate a plurality of acoustic path parameters in accordance with at least one of prior acoustic path parameters, prior far-end features, and prior near-end features; and

generate a plurality of error features based on the far-end features, a plurality of microphone features, and the acoustic path parameters.

13. The system of claim 1 , wherein:

the far-end signal supplied to the contextual-attention neural network comprises a current frame of far-end features and a causal window of a plurality of previous frames of far-end features; and

the near-end signal supplied to the contextual-attention neural network comprises a current frame of near-end features and the causal window of a plurality of previous frames of near-end features.

14. The system of claim 13 , wherein:

the estimated near-end features comprise a current frame of estimated near-end features and the causal window of a plurality of previous frames of estimated near-end features; and

the contextual-attention neural network is trained by iteratively training a plurality of parameters configuring the contextual-attention neural network to minimize differences between the estimated near-end features and a plurality of ground truth near-end features of a plurality of training data.

15. The system of claim 14 , wherein the training data is generated by:

loading a corpus of training data comprising recorded utterances from a plurality of different human speakers;

selecting a plurality of pairs of human speakers; and

for each pair of human speakers:

concatenating multiple utterances of a first speaker of the pair of human speakers to generate a training far-end signal;

transforming the concatenated utterances to simulate an acoustic path to generate a training echo signal;

padding an utterance of a second speaker of the pair of human speakers to generate a training near-end signal having a length equal to the training far-end signal; and

mixing the training echo signal with the training near-end signal to generate a training microphone signal.

16. The system of claim 15 , wherein the transforming the concatenated utterances comprises convolving the far-end signal with a room impulse response of a simulated room.

17. The system of claim 16 , wherein the transforming the concatenated utterances further comprises:

applying hard clipping to the far-end signal to generate a clipped far-end signal; and

applying sigmoidal distortion to the clipped far-end signal.

18. A method comprising:

receiving a far-end signal and a near-end signal comprising at least an echo signal corresponding to the far-end signal;

generating, by a contextual-attention neural network, a plurality of estimated near-end features based at least in part on the received near-end signal and the received far-end signal, without using an estimated echo signal; and

generating, based on the plurality of estimated near-end features, an estimated near-end signal, the estimated near-end signal corresponding to the near-end signal without the echo signal.

19. The method of claim 18 , further comprising:

generating, based on the far-end signal, a plurality of far-end features; and

generating, based on the near-end signal, a plurality of near-end features,

wherein the plurality of estimated near-end features are further generated by inputting one or more near-end features of the plurality of near-end features and one or more far-end features of the plurality of far-end features to the contextual-attention neural network.

20. One or more non-transitory computer-readable storage mediums comprising instructions, which when executed by a processor, cause the processor to:

receive a far-end signal and a near-end signal comprising at least an echo signal corresponding to the far-end signal;

generate, by a contextual-attention neural network, a plurality of estimated near-end features based at least in part on the received near-end signal and the received far-end signal, without using an estimated echo signal; and

generate, based on the plurality of estimated near-end features, an estimated near-end signal, the estimated near-end signal corresponding to the near-end signal without the echo signal.