A New Paradigm for Speaker-Independent Training and Speaker Adaptation

A New Paradigm for Speaker-Independent Training and Speaker Adaptation Francis Kubala and Richard Schwartz BBN Systems and Technologies Corporation Cambridge, MA 02138 Abstract This paper reports on two contributions to large vocabu- lary continuous speech recognition. First, we present a new paradigm for speaker-independent (SI) training of hid- den Markov models (HMM), which uses a large amount of speech from a few speakers instead of the traditional prac- tice of using a little speech from many speakers. In addition, combination of the training speakers is done by averaging the statistics of independently trained models rather than the usual pooling of all the speech data from many speakers prior to training. With only 12 training speakers for SI recogni- tion, we achieved a 7.5% word error rate on a standard grammar and test set from the DARPA Resource Manage- ment corpus. This performance is comparable to our best condition for this test suite, using 109 training speakers. Second, we show a significant improvement for speaker adaptation (SA) using the new SI corpus and a small amount of speech from the new (target) speaker. A probabilistic spectral mapping is estimated independently for each train- ing (reference) speaker and the target speaker. Each refer- ence model is transformed to the space of the target speaker and combined by averaging. Using only 40 utterances from the target speaker for adaptation, the error rate dropped to 4.1% -- a 45% reduction in error compared to the SI result. 1. Introduction One important scenario for the use of spoken language sys- tems (SLS) by new speakers is to start with a SI corpus or model and have the system adapt as the new users interact with the system. Once the interaction has begun, the sys- tem has the opportunity to collect speaker-dependent data of known orthographic transcription from the target speaker. After a small sample of speech has been collected, the sys- tem should be able to adapt so as to significantly increase performance compared to the original SI model. The success of this scenario depends on the adaptation being powerful enough to generalize from a small sample of speaker-specific speech in which most of the phonetic contexts of the lan- guage are not observed. Furthermore, it depends on having a SI speech corpus which is amenable to speaker adaptation. It is a widely held belief that speech used for training SI models must be collected from many speakers. It is also commonly accepted that collecting only a small sample of speech from each training speaker is a reasonable compro- mise to make in the effort to collect as many speakers as possible. While this compromise may be reasonable for SI recognition, several efforts to use such a corpus as a ba- sis for speaker adaptation have failed to make significant improvements. Recently, we have discovered that adequate SI perfor- mance can be achieved with far less speaker coverage than conventionally thought necessary, but with much better sam- piing of each training speaker's speech. Specifically, we …

Content truncated to 500 words