Speech recognition method, device, and computer-readable storage medium

1. A speech recognition method, comprising:

obtaining a first voice signal collected from a first microphone in a microphone array and a second voice signal collected from a second microphone in the microphone array, the second microphone being different from the first microphone;

extracting enhanced features associated with the first voice signal and the second voice signal through a neural network;

obtaining a speech recognition result based on the enhanced features;

determining, based on the enhanced features, whether the first voice signal and the second voice signal involve a wakeup word; and

in response to determining that the first voice signal and the second voice signal involve the wakeup word, initiating a character recognition process,

wherein obtaining the speech recognition result comprises:

compressing the enhanced features associated with the first voice signal and the second voice signal respectively into a dual-channel with a predetermined bandwidth size equivalent to a bandwidth of a single-channel of the enhanced features associated with one of the first or second voice signals; and

providing the compressed enhanced features to a streaming multi-layer truncated attention model.

2. The method of claim 1 , wherein extracting the enhanced features associated with the first voice signal and the second voice signal through the neural network comprises:

performing complex Fourier transform on the first voice signal and the second voice signal, respectively, to obtain a transformed first voice signal and a transformed second voice signal;

performing complex convolution, complex offset, and complex linear transformation on the transformed first voice signal and the transformed second voice signal through a complex convolutional neural network to obtain complex features; and

converting the complex features into enhanced features in real number.

3. The method of claim 2 , wherein obtaining the speech recognition result comprises:

determining, based on the enhanced features, a character output corresponding to the first voice signal and the second voice signal through the streaming multi-layer truncated attention model.

4. The method of claim 2 , further comprising:

determining a direction of a target sound source associated with the first voice signal and the second voice signal based on the enhanced features; and

turning on a reminder light associated with the direction determined.

5. The method of claim 1 , wherein extracting the enhanced features associated with the first voice signal and the second voice signal through the neural network comprises:

obtaining a third voice signal collected from a third microphone in the microphone array; and

extracting enhanced features associated with the first voice signal, the second voice signal and the third voice signal through the neural network.

6. The method of claim 1 , further comprising:

obtaining a same number of multi-channel far-field voice signals as microphones in the microphone array, the multi-channel far-field voice signals at least comprising a first far-field voice signal and a second far-field voice signal; and

training an integrated end-to-end speech enhancement and recognition model by using the multi-channel far-field voice signals.

7. The method of claim 6 , wherein obtaining the same number of multi-channel far-field voice signals as the microphones in the microphone array comprise:

simulating, based on near-field voice signals, the multi-channel far-field voice signals in real time through a random noise addition.

8. The method of claim 7 , wherein simulating the multi-channel far-field voice signals in real time through the random noise addition comprises:

randomly setting simulation parameters: configurations of a room, a position of the microphone array in the room, a position of a target sound source in the room, and a position of a noise source in the room, the configurations of the room comprising a length, a width, and a height of the room, and a wall reflection coefficient.

9. The method of claim 8 , wherein simulating the multi-channel far-field voice signals in real time through the random noise addition comprises:

generating, based on the simulation parameters, a first group of impulse responses for the near-field voice signals and a second group of impulse responses for noise signals randomly selected.

10. The method of claim 9 , wherein simulating the multi-channel far-field voice signals in real time through the random noise addition comprises:

generating the multi-channel far-field voice signals based on the near-field voice signals, the first group of impulse responses, the noise signals, the second group of impulse responses, and a signal-to-noise ratio.

11. An electronic device, comprising:

one or more processors; and

a storage device configured to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the electronic device is caused to implement a speech recognition method, the method comprising:

obtaining a first voice signal collected from a first microphone in a microphone array and a second voice signal collected from a second microphone in the microphone array, the second microphone being different from the first microphone;

extracting enhanced features associated with the first voice signal and the second voice signal through a neural network;

obtaining a speech recognition result based on the enhanced features;

determining, based on the enhanced features, whether the first voice signal and the second voice signal involve a wakeup word; and

in response to determining that the first voice signal and the second voice signal involve the wakeup word, initiating a character recognition process,

wherein obtaining the speech recognition result comprises:

compressing the enhanced features associated with the first voice signal and the second voice signal respectively into a dual-channel with a predetermined bandwidth size equivalent to a bandwidth of a single-channel of the enhanced features associated with one of the first or second voice signals; and

providing the compressed enhanced features to a streaming multi-layer truncated attention model.

12. The electronic device of claim 11 , wherein extracting the enhanced features associated with the first voice signal and the second voice signal through the neural network comprises:

performing complex Fourier transform on the first voice signal and the second voice signal, respectively, to obtain a transformed first voice signal and a transformed second voice signal;

performing complex convolution, complex offset, and complex linear transformation on the transformed first voice signal and the transformed second voice signal through a complex convolutional neural network to obtain complex features; and

converting the complex features into enhanced features in real number.

13. The electronic device of claim 12 , wherein obtaining the speech recognition result comprises:

determining, based on the enhanced features, a character output corresponding to the first voice signal and the second voice signal through the streaming multi-layer truncated attention model.

14. The electronic device of claim 12 , wherein the method further comprises:

determining a direction of a target sound source associated with the first voice signal and the second voice signal based on the enhanced features; and

turning on a reminder light associated with the direction determined.

15. The electronic device of claim 11 , wherein extracting the enhanced features associated with the first voice signal and the second voice signal through the neural network comprises:

obtaining a third voice signal collected from a third microphone in the microphone array; and

extracting enhanced features associated with the first voice signal, the second voice signal and the third voice signal through the neural network.

16. A computer-readable storage medium having a computer program stored thereon, wherein when the program is executed by a processor, a speech recognition method is implemented, the method comprising:

obtaining a first voice signal collected from a first microphone in a microphone array and a second voice signal collected from a second microphone in the microphone array, the second microphone being different from the first microphone;

extracting enhanced features associated with the first voice signal and the second voice signal through a neural network;

obtaining a speech recognition result based on the enhanced features;

determining, based on the enhanced features, whether the first voice signal and the second voice signal involve a wakeup word; and

in response to determining that the first voice signal and the second voice signal involve the wakeup word, initiating a character recognition process,

wherein obtaining the speech recognition result comprises:

compressing the enhanced features associated with the first voice signal and the second voice signal respectively into a dual-channel with a predetermined bandwidth size equivalent to a bandwidth of a single-channel of the enhanced features associated with one of the first or second voice signals; and

providing the compressed enhanced features to a streaming multi-layer truncated attention model.