Evaluating Rapid Unsupervised Speaker Adaptation Using Linear Interpolation of HMM-Sufficient Statistics

摘要

Speaker adaptation techniques minimize the effect of speaker variability. It is neccessary to carry out speaker adaptation rapidly using a minimum amount of adaptation data in real-time application. We propose to improve the unsupervised speaker adaptation based on HMM-Sufficient Statistics using linear interpolation. This adaptation technique uses a single arbitrary utterance to provide data for adaptation by means of selecting N-best speakers' Sufficient Statistics. Reducing the selected N-best speakers implies reduction in adaptation time. However, recognition performance is degraded clue to insufficiency of data needed to robustly adapt the model. We introduce linear interpolation of the global HMM-Sufficient Statistics to offset the negative effect of reducing N-best. We achieved a 50 reduction in adaptation time without recognition performance degradation. In our experiment, we have reduced the adaptation time from 10 sec to 5 sec without degrading the recognition performance. Furthermore we compared our method with Vocal Tract Length Normalization (VTLN), Maximum A Posteriori (MAP) and Maximum Likelihood Linear Regression. Moreover, we tested the performance of our approach in office, car, crowd and booth noise environments in 10 dB, 15 dB, 20 dB and 25 dB SNRs.