Speech model refinement with transcription error detection
Reliable transcription error-checking algorithm that uses a word confidence score and a word duration probability to detect transcription errors for improved results through the automatic detection of transcription errors in a corpus. The transcription error-checking algorithm is combined model training so as to use a current model to detect transcription errors, remove utterances which contain incorrect transcription (or manually fix the found errors), and retrain the model. This process can be repeated for several iterations to obtain an improved speech recognition model. The speech model is employed to achieve speech-transcription alignment to obtain a word boundary. Speech recognizer is then utilized to generate a word-lattice. Using the word boundary and word lattice, error detection is computed using a word confidence score and a word duration probability.
1. A computer-implemented error checking system, comprising:
a duration component for computing word duration probability data based on a speech model, and a corpus of transcription data and wave data;
a confidence component for computing a confidence score based on recognition processing of the speech model and wave data, and alignment output of the transcription data and the wave data;
an error component for detecting error based on the duration probability data and the confidence score; and
a training component for retraining the speech model based on removal of the error from the corpus.
2. The system of claim 1 , further comprising a correction component for removing utterance error from the speech model in a first iteration and feeding back the speech model of the first iteration for retraining of the speech model in a second iteration.
3. The system of claim 1 , further comprising a correction component that facilitates manually checking a transcription error and correcting utterance errors for feedback during retraining of the speech model.
4. The system of claim 1 , further comprising an alignment component that forces alignment of the transcription data, the wave data, and the speech model to generate the word duration probability data.
5. The system of claim 1 , further comprising a recognition component for performing the recognition processing on the wave data and the speech model using a word loop grammar and generating a word lattice graph.
6. The system of claim 1 , further comprising a correction component for using transcription error to remove from the corpus utterances having utterance errors.
7. The system of claim 1 , wherein the error component facilitates manually checking a transcription error found automatically by an error checking system.
8. The system of claim 1 , wherein the corpus with transcription error removed is used for retraining the speech model.
9. The system of claim 1 , wherein the confidence score is based on a posterior probability, generalized posterior probability, or a generalized phone posterior probability.
10. The system of claim 1 , further comprising a threshold component for detecting transcription error based on the word duration probability data relative to a threshold value.
11. The system of claim 1 , further comprising a threshold component for detecting transcription error based on the confidence score relative to a threshold value.
12. A computer-implemented method of error checking, comprising:
computing duration probability data by forcing alignment of a speech model and a corpus of transcription data and wave data;
computing a confidence score based on recognition processing;
detecting errors based on the duration probability data and the confidence score; and
retraining the speech model based on removal of the errors from the corpus.
13. The method of claim 12 , further comprising retraining the corpus based on detection of the errors.
14. The method of claim 12 , further comprising calculating the duration probability data based on a word boundary.
15. The method of claim 12 , further comprising generating a word lattice based on recognition processing of the speech model and the wave data, the word lattice employed to compute the word confidence score.
16. The method of claim 12 , further comprising building a histogram of duration distributions for word models.
17. The method of claim 16 , further comprising selecting outlying duration distributions as having the errors.
18. The method of claim 12 , further comprising modeling the duration probability data according to a Gaussian distribution or a Gamma distribution.
19. The method of claim 12 , further comprising detecting error based on silence confidence scores.
20. A computer-implemented system, comprising:
computer-implemented means for computing duration probability data by forcing alignment of a speech model and a corpus of transcription data and wave data;
computer-implemented means for computing a confidence score based on recognition processing;
computer-implemented means for detecting errors based on the duration probability data and the confidence score; and
computer-implemented means for retraining the speech model based on removal of the errors from the corpus.