Generating a task-adapted acoustic model from one or more different corpora
View Patent ↗The present invention generates a task-dependent acoustic model from a supervised task-independent corpus and further adapted it with an unsupervised task dependent corpus. The task-independent corpus includes task-independent training data which has an acoustic representation of words and a sequence of transcribed words corresponding to the acoustic representation. A relevance measure is defined for each of the words in the task-independent data. The relevance measure is used to weight the data associated with each of the words in the task-independent training data. The task-dependent acoustic model is then trained based on the weighted data for the words in the task-independent training data.
1. A method of generating a task-dependent acoustic model from a task-independent (TI) training corpus that includes an acoustic representation of an utterance and a sequence of transcribed words corresponding to the acoustic representation, the method comprising:
deriving a task relevance measure for each word in the TI training corpus, indicative of a relevance of the words to a task; and
generating a task-dependent (TD) acoustic model (AM) based on the TI training corpus and the task relevance measures for the words in the TI training corpus, by training a task-independent (TI) AM based on the TI training corpus, the TI AM including words from the TI training corpus and associated AM parameters, and weighting the AM parameters with the task relevance measures for the words corresponding to the AM parameters.
2. The method of claim 1 wherein generating the TD AM comprises:
combining the weighted AM parameters to obtain the TD AM.
3. The method of claim 1 wherein generating a TD AM comprises:
weighting words in the TI training corpus.
4. The method of claim 3 wherein generating a TD AM comprises:
training the TD AM using the weighted words.
5. The method of claim 1 wherein generating a TD AM comprises:
extracting relevant data from the TI training corpus based on the task relevance measures.
6. The method of claim 5 wherein generating a TD AM comprises:
training the TD AM based on the relevant data.
7. The method of claim 1 , wherein deriving a task relevance measure comprises:
selecting a word from the TI training corpus; and
defining the task relevance measure for the selected word based on a portion of the selected word that is found in the task.
8. The method of claim 7 wherein defining the task relevance measure for the selected word, comprises:
defining the task relevance measure for the selected word based on a number of relevant triphones in the selected word, the number of relevant triphones being triphones in the selected word that are found in the task.
9. The method of claim 8 wherein the relevance measure for the selected word is defined as a ratio of the number of relevant triphones to a total number of phones in the selected word.
10. The method of claim 1 wherein deriving a relevance measure comprises;
selecting a word from the TI training corpus; and
determining whether the entire selected word is in the task.
11. The method of claim 10 wherein deriving a relevance measure comprises:
defining the relevance measure as relevant if the entire selected word is in the task; and
defining the relevance measure as irrelevant if the entire selected word is not in the task.
12. A system for generating a task-dependent (TD) acoustic model (AM) from a task-independent (TI) training corpus, comprising:
a task relevance generator receiving a task input indicative of words relevant to a task and configured to generate a relevance measure for each word in the TI training corpus based on the task input, wherein the task relevance generator is configured to generate the relevance measure for a selected word based on whether the entire selected word is in the task input; and
an AM generator, coupled to the TI training corpus and the task relevance generator and configured to generate the TD AM based on the TI training corpus and the relevance measure.
13. The system of claim 12 wherein the task relevance generator is configured to generate the relevance measure for a selected word based on a number of phonetic units in the selected word that are in the task input.
14. The system of claim 12 wherein the AM generator comprises:
an AM training component configured to train a TI AM having AM parameters associated with words;
a weighting component configured to weight the AM parameters with the relevance measures for the words associated with the AM parameters; and
a parameter combining component configured to combine the weighted AM parameters to obtain the TD AM.
15. The system of claim 12 wherein the words in the TI training corpus are weighted with the relevance measures and wherein the AM generator comprises:
an AM generator generating the TD AM from the weighted words.
16. A computer readable medium storing instructions which, when executed, cause a computer to perform the steps of:
defining a plurality of task relevance measures, each corresponding to a word in a task-independent (TI) training corpus, the task relevance measures each being indicative of a relevance of its corresponding word to a predetermined task; and
generating a task-dependent (TD) acoustic model (AM) based on the TI training corpus and the relevance measures, by generating a TI AM from the TI training corpus, and modifying the TI AM with the relevance measures to obtain the TD AM.
17. The computer readable medium of claim 16 wherein generating the TD AM comprises:
modifying the TI training corpus with the relevance measures; and
generating the TD AM based on the modified TI training corpus.
18. A method of generating a task-dependent acoustic model from a task-independent (TI) training corpus that includes an acoustic representation of an utterance and a sequence of transcribed words corresponding to the acoustic representation, the method comprising:
deriving a task relevance measure for each word in the TI training corpus, indicative of a relevance of the words to a task by selecting a word from the TI training corpus, defining the task relevance measure for the word based on a number of relevant triphones in the selected word, the number of relevant triphones being triphones in the selected word that are found in the task, wherein the relevance measure for the selected word is defined as a ratio of the number of relevant triphones to a total number of phones in the selected word; and
generating a task-dependent (TD) acoustic model (AM) based on the TI training corpus and the task relevance measures for the words in the TI training corpus.
19. A system for generating a task-dependent (TD) acoustic model (AM) from a task-independent (TI) training corpus, comprising:
a task relevance generator receiving a task input indicative of words relevant to a task and configured to generate a relevance measure for each word in the TI training corpus based on the task input; and
an AM generator, coupled to the TI training corpus and the task relevance generator and configured to generate the TD AM based on the TI training corpus and the relevance measure, wherein the words in the TI training corpus are weighted with the relevance measures and wherein the AM generator is configured to generate the TD AM from the weighted words.
20. A computer readable medium storing instructions which, when executed, cause a computer to perform the steps of:
defining a plurality of task relevance measures, each corresponding to a word in a task-independent (TI) training corpus, the task relevance measures each being indicative of a relevance of its corresponding word to a predetermined task; and
generating a task-dependent (TD) acoustic model (AM) based on the TI training corpus and the relevance measures, by modifying the TI training corpus with the relevance measures, and generating the TD AM based on the modified TI training corpus.