Advancing word-based speech recognition processing
Non-limiting examples of the present disclosure describe advancements in acoustic-to-word modeling that improve accuracy in speech recognition processing through the replacement of out-of-vocabulary (OOV) tokens. During the decoding of speech signals, better accuracy in speech recognition processing is achieved through training and implementation of multiple different solutions that present enhanced speech recognition models. In one example, a hybrid neural network model for speech recognition processing combines a word-based neural network model as a primary model and a character-based neural network model as an auxiliary model. The primary word-based model emits a word sequence, and an output of character-based auxiliary model is consulted at a segment where the word-based model emits an OOV token. In another example, a mixed unit speech recognition model is developed and trained to generate a mixed word and character sequence during decoding of a speech signal without requiring generation of OOV tokens.
1. A method comprising:
detecting that a speech signal comprises an out-of-vocabulary (OOV) token using a word-based language model;
propagating the speech signal to a character-based language model for character-based evaluation only when the OOV token is detected in the speech signal by the word-based language model;
generating a character sequence for the OOV token; and
outputting a speech recognition result for the speech signal that comprises the generated character sequence for the OOV token.
2. The method of claim 1 , wherein the speech signal is processed using a hybrid neural network model that comprises an acoustic-to-word model for detection of the OOV token and character-based auxiliary model evaluation for evaluation of the OOV token.
3. The method of claim 2 , further comprising: training the hybrid neural network model based on processing associated with the speech recognition result.
4. The method of claim 3 , wherein the training further comprises receiving, from a productivity service, usage data associated with an interaction with the speech recognition result and updating training data of the hybrid neural network model based on the usage data.
5. The method of claim 2 , wherein the hybrid neural network model is a hybrid Connectionist Temporal Classification (CTC) model that comprises an acoustic-to-word CTC model for evaluation of the speech signal and a character-based CTC model for evaluation of the OOV token.
6. The method of claim 5 , wherein the hybrid neural network model is a hybrid Connectionist Temporal Classification (CTC) model that comprises an acoustic-to-word CTC model is trained to identify frequent words, and wherein the OOV token is generated when the speech signal is identified as an infrequent word that is not recognized by the acoustic-to-word CTC model.
7. The method of claim 1 , further comprising: collapsing the character sequence into an output unit for the OOV token, wherein the outputting outputs the output unit in the speech recognition result.
8. The method of claim 1 , wherein the outputting comprises propagating the speech recognition result to an application or productivity service.
9. A method comprising:
receiving a speech signal;
decoding the speech signal using a mixed unit speech recognition model that is trained based on word and character sequences;
generating, for the speech signal, a mixed word and character sequence based on an evaluation of the speech signal by the mixed unit speech recognition model, wherein the mixed unit speech recognition model applies a word-based language model to detect an out of vocabulary (OOV) token and applies a character-based language model to evaluate the OOV token only when the OOV token is detected in the speech signal by the word-based language model;
decomposing the mixed word and character sequence; and
outputting a speech recognition result for the speech signal that comprises the decomposed mixed word and character sequence.
10. The method of claim 9 , further comprising: collapsing the character sequence of the decomposed mixed word and character sequence, wherein the speech recognition result, output in the outputting, comprises the collapsed character sequence.
11. The method of claim 9 , wherein the mixed unit speech recognition model is a mixed Connectionist Temporal Classification (CTC) model.
12. The method of claim 9 , wherein the speech signal is received during real-time execution of an application or service.
13. The method of claim 9 , wherein the outputting comprises propagating the speech recognition result to an application or service for subsequent processing.
14. The method of claim 9 , further comprising: updating training data for the mixed unit speech recognition model based on usage data, of the speech recognition result, that is associated with an application or service.
15. A system comprising:
at least one processor; and
a memory, operatively connected with the at least one processor, storing computer-executable instructions that, when executed by the at least one processor, causes the at least one processor to execute a method that comprises:
receiving a speech signal;
decoding the speech signal using a mixed unit speech recognition model that is trained based on word and character sequences;
generating, for the speech signal, a mixed word and character sequence based on an evaluation of the speech signal by the mixed unit speech recognition model, wherein the mixed unit speech recognition model applies a word-based language model to detect an out of vocabulary (OOV) token and applies a character-based language model to evaluate the OOV token only when the OOV token is detected in the speech signal by the word-based language model;
decomposing the mixed word and character sequence; and
outputting a speech recognition result for the speech signal that comprises the decomposed mixed word and character sequence.
16. The system of claim 15 , wherein the method, executed by the at least one processor, further comprises: collapsing the character sequence of the decomposed mixed word and character sequence, wherein the speech recognition result, output in the outputting, comprises the collapsed character sequence.
17. The system of claim 15 , wherein the mixed unit speech recognition model is a mixed Connectionist Temporal Classification (CTC) model.
18. The method of claim 15 , wherein the speech signal is received during real-time execution of an application or service.
19. The method of claim 15 , wherein the outputting comprises propagating the speech recognition result to an application or service for subsequent processing.
20. The system of claim 15 , wherein the method, executed by the at least one processor, further comprises: updating training data for the mixed unit speech recognition model based on usage data, of the speech recognition result, that is associated with an application or service.