Speech recognition based on pronunciation modeling
View Patent ↗A system and method for performing speech recognition is disclosed. The method comprises receiving an utterance, applying the utterance to a recognizer with a language model having pronunciation probabilities associated with unique word identifiers for words given their pronunciations and presenting a recognition result for the utterance. Recognition improvement is found by moving a pronunciation model from a dictionary to the langue model.
1. A method comprising:
training a language model by approximating large quantities of transcribed speech using a plurality of speaker dependent phonemic transcription based datasets that are based on pronunciation models obtained for each of a plurality of speakers;
incorporating, in the language model, pronunciation probabilities associated with respective unique labels for each different pronunciation of a word, wherein the respective unique labels for most frequent words and most frequent word pairs indicate a special status in the language model;
removing, from a pronunciation dictionary, the pronunciation probabilities;
applying, via a processor, an utterance to a recognizer with the language model to yield a recognition result; and
presenting the recognition result for the utterance.
2. The method of claim 1 , wherein the language model is generated by modeling word and pronunciation dependencies across word boundaries.
3. The method of claim 1 , wherein at least one of contextual dependencies and consistency in pronunciation style exist throughout the utterance.
4. The method of claim 1 , wherein the pronunciation probabilities in the language model further comprise pronunciation dependent word pairs as lexical items that change the behavior of the language model to approximate higher order n-gram language models.
5. The method of claim 1 , wherein having the pronunciation probabilities in the language model provides for wide context pronunciation modeling which enables the determination of a probability of observing a particular word with a particular pronunciation based on a long-word/pronunciation history.
6. The method of claim 1 , wherein the pronunciation probabilities comprise a set of most frequent words each with more than one pronunciation alternative.
7. The method of claim 6 , wherein the more than one pronunciation alternative is in the language model.
8. A non-transitory computer-readable storage medium having stored therein instructions which, when executed by a processor, cause the processor to perform a method comprising:
training a language model by approximating large quantities of transcribed speech using a plurality of speaker dependent phonemic transcription based datasets that are based on pronunciation models obtained for each of a plurality of speakers;
incorporating, in the language model, pronunciation probabilities associated with respective unique labels for each different pronunciation of a word, wherein the respective unique labels for most frequent words and most frequent word pairs indicate a special status in the language model;
removing, from a pronunciation dictionary, the pronunciation probabilities;
applying an utterance to a recognizer with the language model to yield a recognition result; and
presenting the recognition result for the utterance.
9. The non-transitory computer-readable storage medium of claim 8 , wherein the language model is generated by modeling word and pronunciation dependencies across word boundaries.
10. The non-transitory computer-readable storage medium of claim 8 , wherein at least one of contextual dependencies and consistency in pronunciation style exist throughout the utterance.
11. The non-transitory computer-readable storage medium of claim 8 , wherein the pronunciation probabilities in the language model further comprise pronunciation dependent word pairs as lexical items that change the behavior of the language model to approximate higher order n-gram language models.
12. The non-transitory computer-readable storage medium of claim 8 , wherein having the pronunciation probabilities in the language model provides for wide context pronunciation modeling which enables the determination of a probability of observing a particular word with a particular pronunciation based on a long-word/pronunciation history.
13. The non-transitory computer-readable storage medium of claim 8 , wherein the pronunciation probabilities comprise a set of most frequent words each with more than one pronunciation alternative.
14. The non-transitory computer-readable storage medium of claim 13 , wherein the more than one pronunciation alternative is in the language model.
15. A system comprising:
a processor, and a non-transitory computer-readable medium having stored therein instructions which, when executed by the processor, cause the processor to perform a method comprising:
training a language model by approximating large quantities of transcribed speech using a plurality of speaker dependent phonemic transcription based datasets that are based on pronunciation models obtained for each of a plurality of speakers;
incorporating, in the language model, pronunciation probabilities associated with respective unique labels for each different pronunciation of a word, wherein the respective unique labels for most frequent words and most frequent word pairs indicate a special status in the language model; removing, from a pronunciation dictionary, the pronunciation probabilities;
applying an utterance to a recognizer with the language model to yield a recognition result; and
presenting the recognition result for the utterance.
16. The system of claim 15 , wherein the language model is generated by modeling word and pronunciation dependencies across word boundaries.
17. The system of claim 15 , wherein at least one of contextual dependencies and consistency in pronunciation style exist throughout the utterance.
18. The system of claim 15 , wherein the pronunciation probabilities in the language model further comprise pronunciation dependent word pairs as lexical items that change the behavior of the language model to approximate higher order n-gram language models.