EUROSPEECH 2003 - INTERSPEECH 2003
Extending the phone set is one common approach for dealing with phonetic confusions in spontaneous speech. We propose using likelihood ratio test as a confidence measure for automatic phone set extension to model phonetic confusions. We first extend the standard phone set using dynamic programming (DP) alignment to cover all possible phonetic confusions in training data. Likelihood ratio test is then used as a confidence measure to optimize the extended phonetic units to represent the acoustic samples between two standard phonetic units with high confusability. The optimum set of extended phonetic units is combined with the standard phone set to form a multiple pronunciation dictionary. The effectiveness of this approach is evaluated on spontaneous Mandarin telephony speech. It gives an encouraging 1.09% absolute syllable error rate reduction. Using the extended phone set provides a good balance between the demands of high resolution acoustic model and the available training data.
Bibliographic reference. Liu, Yi / Fung, Pascale (2003): "Automatic phone set extension with confidence measure for spontaneous speech", In EUROSPEECH-2003, 2741-2744.