Method for training speech recognition model, device and storage medium
A method for training a speech recognition model, a device and a storage medium, which relate to the field of computer technologies, and particularly to the fields of speech recognition technologies, deep learning technologies, or the like, are disclosed. The method for training a speech recognition model includes: obtaining a fusion probability of each of at least one candidate text corresponding to a speech based on an acoustic decoding model and a language model; selecting a preset number of one or more candidate texts based on the fusion probability of each of the at least one candidate text, and determining a predicted text based on the preset number of one or more candidate texts; and obtaining a loss function based on the predicted text and a standard text corresponding to the speech, and training the speech recognition model based on the loss function.
1. A method for training a speech recognition model, the speech recognition model comprising an acoustic decoding model and a language model, the method comprising:
obtaining a fusion probability of each of at least one candidate text corresponding to a speech based on the acoustic decoding model and the language model;
selecting a preset number of one or more candidate texts based on the fusion probability of each of the at least one candidate text, and determining a predicted text based on the preset number of one or more candidate texts; and
obtaining a loss function based on the predicted text and a standard text corresponding to the speech, and training the speech recognition model based on the loss function,
wherein the obtaining the loss function based on the predicted text and the standard text corresponding to the speech comprises:
obtaining an accumulated number of errors of the predicted text based on the predicted text and the standard text corresponding to the speech, the accumulated error number being obtained based on a historical error number and a current error number; and
obtaining the loss function based on the accumulated error number of the predicted text.
2. The method according to claim 1 , wherein the obtaining the fusion probability of each of the at least one candidate text corresponding to the speech based on the acoustic decoding model and the language model comprises:
processing an output text at a previous moment and acoustic related information at a current moment using the acoustic decoding model to obtain a first probability corresponding to the at least one candidate text corresponding to the speech;
processing the output text at the previous moment using the language model to obtain a second probability corresponding to the candidate text; and
obtaining the fusion probability of the candidate text based on the first probability and the second probability.
3. The method according to claim 2 , wherein the obtaining the fusion probability of the candidate text based on the first probability and the second probability comprises:
for the candidate text, performing weighted summation of the first probability and the second probability to obtain a weighted summation value, and determining the weighted summation value as the fusion probability of the corresponding candidate text.
4. The method according to claim 1 , wherein the determining the predicted text based on the preset number of the one or more candidate texts comprises:
if the preset number of the one or more candidate texts comprise the standard text, determining the preset number of the one or more candidate texts as the predicted text; if the preset number of the one or more candidate texts do not comprise the standard text, replacing one candidate text of the preset number of the one or more candidate texts with the standard text to obtain one or more texts after the replacing, and determining the one or more texts after the replacing as the predicted text.
5. The method according to claim 1 , wherein the speech recognition model further comprises an encoder and an attention model, the acoustic related information is features after an attention processing operation, and the method further comprises:
extracting acoustic features of the speech;
encoding the acoustic features using the encoder to obtain encoded features; and
performing the attention processing operation on the encoded features using the attention model to obtain the features after the attention processing operation.
6. An electronic device, comprising:
at least one processor; and
a memory connected with the at least one processor communicatively;
wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to carry out a method for training a speech recognition model, which comprises:
obtaining a fusion probability of each of at least one candidate text corresponding to a speech based on an acoustic decoding model and a language model;
selecting a preset number of one or more candidate texts based on the fusion probability of each of the at least one candidate text, and determining a predicted text based on the preset number of one or more candidate texts; and
obtaining a loss function based on the predicted text and a standard text corresponding to the speech, and training the speech recognition model based on the loss function,
wherein the obtaining the loss function based on the predicted text and the standard text corresponding to the speech comprises:
obtaining an accumulated number of errors of the predicted text based on the predicted text and the standard text corresponding to the speech, the accumulated error number being obtained based on a historical error number and a current error number; and
obtaining the loss function based on the accumulated error number of the predicted text.
7. The electronic device according to claim 6 , wherein the obtaining the fusion probability of each of the at least one candidate text corresponding to the speech based on the acoustic decoding model and the language model comprises:
processing an output text at a previous moment and acoustic related information at a current moment using the acoustic decoding model to obtain a first probability corresponding to the at least one candidate text corresponding to the speech;
processing the output text at the previous moment using the language model to obtain a second probability corresponding to the candidate text; and
obtaining the fusion probability of the candidate text based on the first probability and the second probability.
8. The electronic device according to claim 7 , wherein the obtaining the fusion probability of the candidate text based on the first probability and the second probability comprises:
for the candidate text, performing weighted summation of the first probability and the second probability to obtain a weighted summation value, and determining the weighted summation value as the fusion probability of the corresponding candidate text.
9. The electronic device according to claim 6 , wherein the determining the predicted text based on the preset number of the one or more candidate texts comprises:
if the preset number of the one or more candidate texts comprise the standard text, determining the preset number of the one or more candidate texts as the predicted text; if the preset number of the one or more candidate texts do not comprise the standard text, replacing one candidate text of the preset number of the one or more candidate texts with the standard text to obtain one or more texts after the replacing, and determining the one or more texts after the replacing as the predicted text.
10. The electronic device according to claim 6 , wherein the speech recognition model further comprises an encoder and an attention model, the acoustic related information is features after an attention processing operation, and the method further comprises:
extracting acoustic features of the speech;
encoding the acoustic features using the encoder to obtain encoded features; and
performing the attention processing operation on the encoded features using the attention model to obtain the features after the attention processing operation.
11. A non-transitory computer readable storage medium comprising computer instructions, which, when executed by a computer, cause the computer to carry out a method for training a speech recognition model, which comprises:
obtaining a fusion probability of each of at least one candidate text corresponding to a speech based on an acoustic decoding model and a language model;
selecting a preset number of one or more candidate texts based on the fusion probability of each of the at least one candidate text, and determining a predicted text based on the preset number of one or more candidate texts; and
obtaining a loss function based on the predicted text and a standard text corresponding to the speech, and training the speech recognition model based on the loss function,
wherein the obtaining the loss function based on the predicted text and the standard text corresponding to the speech comprises:
obtaining an accumulated number of errors of the predicted text based on the predicted text and the standard text corresponding to the speech, the accumulated error number being obtained based on a historical error number and a current error number; and
obtaining the loss function based on the accumulated error number of the predicted text.
12. The non-transitory computer readable storage medium according to claim 11 , wherein the obtaining the fusion probability of each of the at least one candidate text corresponding to the speech based on the acoustic decoding model and the language model comprises:
processing an output text at a previous moment and acoustic related information at a current moment using the acoustic decoding model to obtain a first probability corresponding to the at least one candidate text corresponding to the speech;
processing the output text at the previous moment using the language model to obtain a second probability corresponding to the candidate text; and
obtaining the fusion probability of the candidate text based on the first probability and the second probability.
13. The non-transitory computer readable storage medium according to claim 12 , wherein the obtaining the fusion probability of the candidate text based on the first probability and the second probability comprises:
for the candidate text, performing weighted summation of the first probability and the second probability to obtain a weighted summation value, and determining the weighted summation value as the fusion probability of the corresponding candidate text.
14. The non-transitory computer readable storage medium according to claim 11 , wherein the determining the predicted text based on the preset number of the one or more candidate texts comprises:
if the preset number of the one or more candidate texts comprise the standard text, determining the preset number of the one or more candidate texts as the predicted text; if the preset number of the one or more candidate texts do not comprise the standard text, replacing one candidate text of the preset number of the one or more candidate texts with the standard text to obtain one or more texts after the replacing, and determining the one or more texts after the replacing as the predicted text.
15. The non-transitory computer readable storage medium according to claim 11 , wherein the speech recognition model further comprises an encoder and an attention model, the acoustic related information is features after an attention processing operation, and the method further comprises:
extracting acoustic features of the speech;
encoding the acoustic features using the encoder to obtain encoded features; and
performing the attention processing operation on the encoded features using the attention model to obtain the features after the attention processing operation.