Method and device extracting acoustic feature based on convolution neural network and terminal device

1. A method for extracting an acoustic feature based on a convolution neural network, comprising:

arranging speech to be recognized into a speech spectrogram with a predetermined dimension number; and

recognizing the speech spectrogram with the predetermined dimension number by the convolution neural network, to obtain the acoustic feature of the speech to be recognized,

wherein the convolution neural network comprises a structure of one of the following:

a residual network structure; and

a jump link structure;

wherein before recognizing the speech spectrogram with the predetermined dimension number by the convolution neural network, further comprising:

configuring a model of the structure of the convolution neural network;

wherein configuring the model of the structure of the convolution neural network comprises:

for a 64-channel filter block consisting of a convolution directed acycline graph of a 64-channel filter bank, performing a down-sampling by a pooling layer both in a time domain and in a frequency domain;

for a 128-channel filter block consisting of a convolution directed acycline graph of a 128-channel filter bank, performing the down-sampling by the pooling layer both in the time domain and in the frequency domain;

for a 256-channel filter block consisting of a convolution directed acycline graph of a 256-channel filter bank, performing the down-sampling by the pooling layer in the frequency domain; and

for a 512-channel filter block consisting of a convolution directed acycline graph of a 512-channel filter bank, performing the down-sampling by the pooling layer in the frequency domain.

2. The method according to claim 1 , wherein arranging speech to be recognized into a speech spectrogram with a predetermined dimension number comprises:

extracting a predetermined multidimensional feature vector from the speech to be recognized every predetermined time interval to arrange the speech to be recognized into the speech spectrogram with the predetermined dimension number.

3. A device for extracting an acoustic feature based on a convolution neural network, comprising:

one or more processors;

a memory, configured to store one or more program modules executable by the one or more processors, wherein the one or more program modules comprise:

a generating module, configured to arrange speech to be recognized into a speech spectrogram with a predetermined dimension number; and

a recognizing module, configured to recognize the speech spectrogram with the predetermined dimension number by the convolution neural network to obtain the acoustic feature of the speech to be recognized,

wherein the convolution neural network comprises a structure of one of the following:

a residual network structure; and

a jump link structure;

wherein the one or more program modules further comprise:

a configuring module, configured to configure a model of the structure of the convolution neural network before the recognizing module recognizes the speech spectrogram with the predetermined dimension number;

wherein, the configuring module is configured to:

for a 64-channel filter block consisting of a convolution directed acycline graph of a 64-channel filter bank, perform a down-sampling by a pooling layer both in a time domain and in a frequency domain;

for a 128-channel filter block consisting of a convolution directed acycline graph of a 128-channel filter bank, perform the down-sampling by the pooling layer both in the time domain and in the frequency domain;

for a 256-channel filter block consisting of a convolution directed acycline graph of a 256-channel filter bank, perform the down-sampling by the pooling layer in the frequency domain; and

for a 512-channel filter block consisting of a convolution directed acycline graph of a 512-channel filter bank, perform the down-sampling by the pooling layer in the frequency domain.

4. The device according to claim 3 , wherein,

the generating module is configured to extract a predetermined multidimensional feature vector from the speech to be recognized every predetermined time interval to arrange the speech to be recognized into the speech spectrogram with the predetermined dimension number.

5. A non-transitory computer readable storage medium comprising computer executable instructions configured to perform a method for extracting an acoustic feature based on a convolution neural network when executed by a computer processor, the method comprising:

arranging speech to be recognized into a speech spectrogram with a predetermined dimension number; and

recognizing the speech spectrogram with the predetermined dimension number by the convolution neural network, to obtain the acoustic feature of the speech to be recognized,

wherein the convolution neural network comprises a structure of one of the following:

a residual network structure; and

a jump link structure;

wherein before recognizing the speech spectrogram with the predetermined dimension number by the convolution neural network, further comprising:

configuring a model of the structure of the convolution neural network;

wherein configuring the model of the structure of the convolution neural network comprises:

for a 64-channel filter block consisting of a convolution directed acycline graph of a 64-channel filter bank, performing a down-sampling by a pooling layer both in a time domain and in a frequency domain;

for a 128-channel filter block consisting of a convolution directed acycline graph of a 128-channel filter bank, performing the down-sampling by the pooling layer both in the time domain and in the frequency domain;

for a 256-channel filter block consisting of a convolution directed acycline graph of a 256-channel filter bank, performing the down-sampling by the pooling layer in the frequency domain; and

for a 512-channel filter block consisting of a convolution directed acycline graph of a 512-channel filter bank, performing the down-sampling by the pooling layer in the frequency domain.