Speech processing method, apparatus, electronic device, and computer-readable storage medium

1. A speech processing method, performed by an electronic device, the method comprising:

determining a first speech feature and a first text bottleneck feature based on to-be-processed speech information;

determining a first combined feature vector based on the first speech feature and the first text bottleneck feature, wherein the determining the first text bottleneck feature based on the to-be-processed speech information comprises:

extracting an N-dimensional filter-bank feature and an M-dimensional pitch feature from the to-be-processed speech information, wherein N and M are positive integers,

splicing the N-dimensional filter-bank feature and the M-dimensional pitch feature to obtain a second speech feature,

inputting the second speech feature into a trained automatic speech recognition (ASR) network, and

extracting the first text bottleneck feature from a linear layer of a bottleneck of the trained ASR network;

inputting the first combined feature vector to a trained unidirectional long short-term memory (LSTM) model;

performing speech processing on the first combined feature vector to obtain speech information after noise reduction; and

transmitting the obtained speech information after noise reduction to another electronic device for playing.

2. The method of claim 1 , wherein the determining the first speech feature based on the to-be-processed speech information comprises:

performing framing and windowing on the to-be-processed speech information; and

extracting the first speech feature from the to-be-processed speech information obtained after the framing and windowing,

wherein the first speech feature includes at least one of a logarithmic power spectrum feature and a Mel-frequency cepstrum coefficient (MFCC) feature.

3. The method of claim 1 , wherein the ASR network is trained by:

(a) aligning a text annotation included in a corpus with an audio file corresponding to the text annotation by using a Gaussian mixture model (GMM) to obtain a first text feature, the corpus being used for training the ASR network;

(b) extracting an N-dimensional filter-bank feature and an M-dimensional pitch feature from the audio file;

(c) splicing the N-dimensional filter-bank feature and the M-dimensional pitch feature to obtain a third speech feature;

(d) inputting the third speech feature to the ASR network and training the ASR network to obtain a second text feature outputted by an output layer of the ASR network;

(e) determining a value of cross entropy (CE) of the ASR network based on a value of the first text feature and a value of the second text feature; and

repeatedly performing steps (a)-(e) to obtain a trained ASR network when a difference between a first value of CE of the ASR network obtained through training and a second value of CE of the ASR network obtained through training at a previous time is in a first threshold range.

4. The method of claim 3 , wherein the ASR network comprises a deep neural network (DNN) with four hidden layers as an input layer, a linear layer of a bottleneck, and a probability distribution softmax layer as an output layer.

5. The method of claim 1 , wherein the performing speech processing on the first combined feature vector to obtain speech information after noise reduction comprises:

performing speech enhancement on the first combined feature vector by using the trained unidirectional LSTM model;

performing inverse feature transformation on a processing result; and

converting speech information from a frequency domain to a time domain to obtain the speech information after noise reduction.

6. The method of claim 1 , wherein the unidirectional LSTM model is trained by:

acquiring speech with noise and speech without noise included in a noise reduction training corpus;

extracting a fourth speech feature and a second text bottleneck feature from the speech with noise;

extracting a fifth speech feature from the speech without noise;

combining the fourth speech feature and the second text bottleneck feature to obtain a second combined feature vector;

inputting the second combined feature vector to the unidirectional LSTM model; and

training the unidirectional LSTM model to obtain a trained unidirectional LSTM model when a minimum mean square error between a reference value outputted by the unidirectional LSTM model and a value of the fifth speech feature is less than or equal to a second threshold.

7. A speech processing apparatus, comprising:

at least one memory configured to store computer program code; and

at least one processor configured to access said computer program code and operate as instructed by said computer program code, said computer program code including:

first determining code configured to cause the at least one processor to determine a first speech feature and a first text bottleneck feature based on to-be-processed speech information;

second determining code configured to cause the at least one processor to determine a first combined feature vector based on the first speech feature and the first text bottleneck feature, wherein to determine the first text bottleneck feature based on the to-be-processed speech information, the second determining code is further configured to cause the at least one processor to:

extract an N-dimensional filter-bank feature and an M-dimensional pitch feature from the to-be-processed speech information, wherein N and M are positive integers,

splice the N-dimensional filter-bank feature and the M-dimensional pitch feature to obtain a second speech feature,

input the second speech feature to a trained automatic speech recognition (ASR) network, and

extract the first text bottleneck feature from a linear layer of a bottleneck of the trained ASR network;

inputting code configured to cause the at least one processor to input the first combined feature vector to a trained unidirectional long short-term memory (LSTM) model;

performing code configured to cause the at least one processor to perform speech processing on the first combined feature vector to obtain speech information after noise reduction; and

transmitting code configured to cause the at least one processor to transmit the speech information after noise reduction to another electronic device for playing.

8. The apparatus of claim 7 , wherein the second determining code is further configured to cause the at least one processor to:

perform framing and windowing on the to-be-processed speech information; and

extract the first speech feature from the to-be-processed speech information obtained after the framing and windowing,

wherein the first speech feature includes at least one of a logarithmic power spectrum feature and a Mel-frequency cepstrum coefficient (MFCC) feature.

9. The apparatus of claim 7 , wherein the ASR network is trained by:

(a) aligning a text annotation included in a corpus with an audio file corresponding to the text annotation by using a Gaussian mixture model (GMM) to obtain a first text feature, the corpus being used for training the ASR network;

(b) extracting an N-dimensional filter-bank feature and an M-dimensional pitch feature from the audio file;

(c) splicing the N-dimensional filter-bank feature and the M-dimensional pitch feature to obtain a third speech feature;

(d) inputting the third speech feature to the ASR network and training the ASR network to obtain a second text feature outputted by an output layer of the ASR network;

(e) determining a value of cross entropy (CE) of the ASR network based on a value of the first text feature and a value of the second text feature; and

repeatedly performing steps (a)-(e) to obtain a trained ASR network when a difference between a first value of CE of the ASR network obtained through training and a second value of CE of the ASR network obtained through training at a previous time is in a first threshold range.

10. The apparatus of claim 7 , wherein the inputting code is further configured to cause the at least one processor to:

perform speech enhancement on the first combined feature vector by using the trained unidirectional LSTM model;

perform inverse feature transformation on a processing result; and

convert speech information from a frequency domain to a time domain to obtain the speech information after noise reduction.

11. The apparatus of claim 7 , wherein the unidirectional LSTM model is trained by:

acquiring speech with noise and speech without noise included in a noise reduction training corpus;

extracting a fourth speech feature and a second text bottleneck feature from the speech with noise;

extracting a fifth speech feature from the speech without noise;

combining the fourth speech feature and the second text bottleneck feature to obtain a second combined feature vector;

inputting the second combined feature vector to the unidirectional LSTM model; and

training the unidirectional LSTM model to obtain a trained unidirectional LSTM model when a minimum mean square error between a reference value outputted by the unidirectional LSTM model and a value of the fifth speech feature is less than or equal to a second threshold.

12. A non-transitory computer-readable storage medium storing computer instructions that, when executed by at least one processor of a device, cause the at least one processor to:

determine a first speech feature and a first text bottleneck feature based on to-be-processed speech information;

determine a first combined feature vector based on the first speech feature and the first text bottleneck feature, wherein the at least one processor is further configured to determine the first text bottleneck feature based on the to-be-processed speech information by:

extracting an N-dimensional filter-bank feature and an M-dimensional pitch feature from the to-be-processed speech information, wherein N and M are positive integers;

splicing the N-dimensional filter-bank feature and the M-dimensional pitch feature to obtain a second speech feature;

inputting the second speech feature into a trained automatic speech recognition (ASR) network; and

extracting the first text bottleneck feature from a linear layer of a bottleneck of the trained ASR network;

input the first combined feature vector to a trained unidirectional long short-term memory (LSTM) model;

perform speech processing on the first combined feature vector to obtain speech information after noise reduction; and

transmit the obtained speech information after noise reduction to another electronic device for playing.

13. The non-transitory computer-readable storage medium of claim 12 , wherein the instructions, when executed by at least one processor of a device, further cause the at least one processor to determine the first speech feature based on the to-be-processed speech information by:

performing framing and windowing on the to-be-processed speech information; and

extracting the first speech feature from the to-be-processed speech information obtained after the framing and windowing,

wherein the first speech feature includes at least one of a logarithmic power spectrum feature and a Mel-frequency cepstrum coefficient (MFCC) feature.

14. The non-transitory computer-readable storage medium of claim 12 , wherein the ASR network is trained by:

(a) aligning a text annotation included in a corpus with an audio file corresponding to the text annotation by using a Gaussian mixture model (GMM) to obtain a first text feature, the corpus being used for training the ASR network;

(b) extracting an N-dimensional filter-bank feature and an M-dimensional pitch feature from the audio file;

(c) splicing the N-dimensional filter-bank feature and the M-dimensional pitch feature to obtain a third speech feature;

(d) inputting the third speech feature to the ASR network and training the ASR network to obtain a second text feature outputted by an output layer of the ASR network;

(e) determining a value of cross entropy (CE) of the ASR network based on a value of the first text feature and a value of the second text feature; and

repeatedly performing steps (a)-(e) to obtain a trained ASR network when a difference between a first value of CE of the ASR network obtained through training and a second value of CE of the ASR network obtained through training at a previous time is in a first threshold range.

15. The non-transitory computer-readable storage medium of claim 14 , wherein the ASR network comprises a deep neural network (DNN) with four hidden layers as an input layer, a linear layer of a bottleneck, and a probability distribution softmax layer as an output layer.

16. The non-transitory computer-readable storage medium of claim 12 , wherein the instructions, when executed by at least one processor of a device, further cause the at least one processor to perform speech processing on the first combined feature vector to obtain speech information after noise reduction by:

performing speech enhancement on the first combined feature vector by using the trained unidirectional LSTM model;

performing inverse feature transformation on a processing result; and

converting speech information from a frequency domain to a time domain to obtain the speech information after noise reduction.

17. The non-transitory computer-readable storage medium of claim 12 , wherein the unidirectional LSTM model is trained by:

acquiring speech with noise and speech without noise included in a noise reduction training corpus;

extracting a fourth speech feature and a second text bottleneck feature from the speech with noise;

extracting a fifth speech feature from the speech without noise;

combining the fourth speech feature and the second text bottleneck feature to obtain a second combined feature vector;

inputting the second combined feature vector to the unidirectional LSTM model; and

training the unidirectional LSTM model to obtain a trained unidirectional LSTM model when a minimum mean square error between a reference value outputted by the unidirectional LSTM model and a value of the fifth speech feature is less than or equal to a second threshold.