Acoustic model training method, speech recognition method, apparatus, device and medium

1. An acoustic model training method, comprising:

performing feature extraction on a training speech signal to obtain an audio feature sequence;

training the audio feature sequence by a phoneme mixed Gaussian Model-Hidden Markov Model to obtain phoneme feature sequence; and

training the phoneme feature sequence by a Deep Neural Net-Hidden Markov Model-sequence training model to obtain a target acoustic model;

wherein training the audio feature sequence by a phoneme mixed Gaussian Model-Hidden Markov Model to obtain a phoneme feature sequence comprises:

training the audio feature sequence by a monophone Mixture Gaussian Model-Hidden Markov Model to obtain the phoneme feature sequence;

training the original monophone mixed Gaussian Model-Hidden Markov Model by the audio feature sequence;

obtaining an original monophone annotation corresponding to each audio feature in the audio feature sequence based on the original monophone Mixture Gaussian Model-Hidden Markov Model;

iteratively training the original monophone Mixture Gaussian Model-Hidden Markov Model based on the audio feature sequence and the original monophone annotation to obtain a target monophone Mixture Gaussian Model-Hidden Markov Model;

aligning each original monophone annotation based on the target monophone Mixture Gaussian Model-Hidden Markov Model to obtain a target monophone feature; and

obtaining a phoneme feature sequence based on the target monophone feature.

2. The acoustic model training method according to claim 1 , wherein performing feature extraction on a training speech signal comprises:

performing pre-emphasis, framing, windowing, endpoint detection, fast Fourier transform, and Mel filter bank and discrete cosine transform on the training speech signal.

3. The acoustic model training method according to claim 1 , wherein training the audio feature sequence by a phoneme mixed Gaussian Model-Hidden Markov Model to obtain a phoneme feature sequence comprises:

training the phoneme feature sequence by a triphone Mixture Gaussian Model-Hidden Markov Model to obtain an updated phoneme feature sequence;

wherein training the phoneme feature sequence by a triphone Mixture Gaussian Model-Hidden Markov Model to obtain an updated phoneme feature sequence comprises:

obtaining an original triphone annotation based on the target monophone feature in the phoneme feature sequence;

inputting the original triphone annotation into the target monophone Mixture Gaussian Model-Hidden Markov Model to obtain an original triphone Mixture Gaussian Model-Hidden Markov Model;

clustering the original triphone Mixture Gaussian Model-Hidden Markov Model to obtain a clustered triphone Mixture Gaussian Model-Hidden Markov Model;

iteratively training the clustered triphone Mixture Gaussian Model-Hidden Markov Model based on the audio feature sequence and the original triphone annotation to obtain a target triphone Mixture Gaussian Model-Hidden Markov Model;

aligning each original triphone annotation based on the target triphone Mixture Gaussian Model-Hidden Markov Model to obtain a target triphone feature; and

obtaining the updated phoneme feature sequence based on the target triphone feature.

4. The acoustic model training method according to claim 3 , wherein training the phoneme feature sequence by a Deep Neural Net-Hidden Markov Model-sequence training model to obtain a target acoustic model comprises:

training an original Deep Neural Net-Hidden Markov Model-sequence training model with the phoneme feature sequence;

obtaining an original state level annotation corresponding to each target monophone feature or each target triphone feature, based on the original Deep Neural Net-Hidden Markov Model-Sequence Training Model;

training the original Deep Neural Net-Hidden Markov Model-Sequence Training Model based on the audio feature sequence and the original state level annotation to obtain an intermediate Deep Neural Net-Hidden Markov Model-Sequence Training Model; and

performing a combined training on the intermediate Deep Neural Net-Hidden Markov Model-sequence training model to obtain the target acoustic model by using a cross-entropy training criterion, a L2-norm training criterion and a Leaky HMM training criterion.

5. A terminal device comprising a memory, a processor, and a computer program stored in the memory and operable on the processor, wherein the processor performs following steps when executing the computer program:

performing feature extraction on a training speech signal to obtain an audio feature sequence;

training the audio feature sequence by a phoneme mixed Gaussian Model-Hidden Markov Model to obtain a phoneme feature sequence; and

training the phoneme feature sequence by a Deep Neural Net-Hidden Markov Model-sequence training model to obtain a target acoustic model;

wherein training the audio feature sequence by a phoneme mixed Gaussian Model-Hidden Markov Model to obtain a phoneme feature sequence comprises:

training the audio feature sequence by a monophone Mixture Gaussian Model-Hidden Markov Model to obtain the phoneme feature sequence;

obtaining an original monophone annotation corresponding to each audio feature in the audio feature sequence based on an original monophone Mixture Gaussian Model-Hidden Markov Model;

iteratively training the original monophone Mixture Gaussian Model-Hidden Markov Model based on the audio feature sequence and the original monophone annotation to obtain a target monophone Mixture Gaussian Model-Hidden Markov Model;

aligning each original monophone annotation based on the target monophone Mixture Gaussian Model-Hidden Markov Model to obtain a target monophone feature; and

obtaining a phoneme feature sequence based on the target monophone feature.

6. The terminal device according to claim 5 , wherein performing feature extraction on a training speech signal comprises:

performing pre-emphasis, framing, windowing, endpoint detection, fast Fourier transform, and Mel filter bank and discrete cosine transform on the training speech signal.

7. The terminal device according to claim 5 , wherein training the audio feature sequence by a phoneme mixed Gaussian Model-Hidden Markov Model to obtain a phoneme feature sequence comprises:

training the phoneme feature sequence by a triphone Mixture Gaussian Model-Hidden Markov Model to obtain an updated phoneme feature sequence;

wherein training the phoneme feature sequence by a triphone Mixture Gaussian Model-Hidden Markov Model to obtain an updated phoneme feature sequence comprises:

obtaining an original triphone annotation based on the target monophone feature in the phoneme feature sequence;

inputting the original triphone annotation into the target monophone Mixture Gaussian Model-Hidden Markov Model to obtain an original triphone Mixture Gaussian Model-Hidden Markov Model;

clustering the original triphone Mixture Gaussian Model-Hidden Markov Model to obtain a clustered triphone Mixture Gaussian Model-Hidden Markov Model;

iteratively training the clustered triphone Mixture Gaussian Model-Hidden Markov Model based on the audio feature sequence and the original triphone annotation to obtain a target triphone Mixture Gaussian Model-Hidden Markov Model;

aligning each original triphone annotation based on the target triphone Mixture Gaussian Model-Hidden Markov Model to obtain a target triphone feature;

obtaining the updated phoneme feature sequence based on the target triphone feature.

8. The terminal device according to claim 7 , wherein training the phoneme feature sequence by a Deep Neural Net-Hidden Markov Model-sequence training model to obtain a target acoustic model comprises:

training an original Deep Neural Net-Hidden Markov Model-sequence training model with the phoneme feature sequence;

obtaining an original state level annotation corresponding to each target monophone feature or each target triphone feature, based on the original Deep Neural Net-Hidden Markov Model-Sequence Training Model;

training the original Deep Neural Net-Hidden Markov Model-Sequence Training Model based on the audio feature sequence and the original state level annotation to obtain an intermediate Deep Neural Net-Hidden Markov Model-Sequence Training Model; and

performing a combined training on the intermediate Deep Neural Net-Hidden Markov Model-sequence training model to obtain the target acoustic model by using a cross-entropy training criterion, a L2-norm training criterion and a Leaky HMM training criterion.

9. A non-transitory computer readable storage medium, the computer readable storage medium stores a computer program, wherein the following steps are performed when the computer program is executed by the processor:

performing feature extraction on a training speech signal to obtain an audio feature sequence;

training the audio feature sequence by a phoneme mixed Gaussian Model-Hidden Markov Model to obtain a phoneme feature sequence; and

training the phoneme feature sequence by a Deep Neural Net-Hidden Markov Model-sequence training model to obtain a target acoustic model;

wherein training the audio feature sequence by a phoneme mixed Gaussian Model-Hidden Markov Model to obtain a phoneme feature sequence comprises:

training the audio feature sequence by a monophone Mixture Gaussian Model-Hidden Markov Model to obtain the phoneme feature sequence;

training the original monophone mixed Gaussian Model-Hidden Markov Model by the audio feature sequence

obtaining an original monophone annotation corresponding to each audio feature in the audio feature sequence based on an original monophone Mixture Gaussian Model-Hidden Markov Model;

iteratively training the original monophone Mixture Gaussian Model-Hidden Markov Model based on the audio feature sequence and the original monophone annotation to obtain a target monophone Mixture Gaussian Model-Hidden Markov Model;

aligning each original monophone annotation based on the target monophone Mixture Gaussian Model-Hidden Markov Model to obtain a target monophone feature; and

obtaining a phoneme feature sequence based on the target monophone feature.

10. The non-transitory computer readable storage medium according to claim 9 , wherein performing feature extraction on a training speech signal comprises:

performing pre-emphasis, framing, windowing, endpoint detection, fast Fourier transform, and Mel filter bank and discrete cosine transform on the training speech signal.

11. The non-transitory computer readable storage medium according to claim 9 , wherein training the audio feature sequence by a phoneme mixed Gaussian Model-Hidden Markov Model to obtain a phoneme feature sequence comprises:

training the phoneme feature sequence by a triphone Mixture Gaussian Model-Hidden Markov Model to obtain an updated phoneme feature sequence;

wherein training the phoneme feature sequence by a triphone Mixture Gaussian Model-Hidden Markov Model to obtain an updated phoneme feature sequence comprises:

obtaining an original triphone annotation based on the target monophone feature in the phoneme feature sequence;

inputting the original triphone annotation into the target monophone Mixture Gaussian Model-Hidden Markov Model to obtain an original triphone Mixture Gaussian Model-Hidden Markov Model;

clustering the original triphone Mixture Gaussian Model-Hidden Markov Model to obtain a clustered triphone Mixture Gaussian Model-Hidden Markov Model;

iteratively training the clustered triphone Mixture Gaussian Model-Hidden Markov Model based on the audio feature sequence and the original triphone annotation to obtain a target triphone Mixture Gaussian Model-Hidden Markov Model;

aligning each original triphone annotation based on the target triphone Mixture Gaussian Model-Hidden Markov Model to obtain a target triphone feature; and

obtaining the updated phoneme feature sequence based on the target triphone feature.

12. The non-transitory computer readable storage medium according to claim 11 , wherein training the phoneme feature sequence by a Deep Neural Net-Hidden Markov Model-sequence training model to obtain a target acoustic model comprises:

training an original Deep Neural Net-Hidden Markov Model-sequence training model with the phoneme feature sequence;

obtaining an original state level annotation corresponding to each target monophone feature or each target triphone feature, based on the original Deep Neural Net-Hidden Markov Model-Sequence Training Model;

training the original Deep Neural Net-Hidden Markov Model-Sequence Training Model based on the audio feature sequence and the original state level annotation to obtain an intermediate Deep Neural Net-Hidden Markov Model-Sequence Training Model; and

performing a combined training on the intermediate Deep Neural Net-Hidden Markov Model-sequence training model to obtain the target acoustic model by using a cross-entropy training criterion, a L2-norm training criterion and a Leaky HMM training criterion.