Third International Conference on Spoken Language Processing (ICSLP 94)
A new method for improving continuous speech recognition (CSR) using speaker adaptation is introduced. The proposed method is unsupervised - no prior information about the speaker is used and there is no training phase involved. The effect of speaker adaptation is achieved by assuming that a set of speaker parameters is constant over the utterance. Both the acoustic information and the speaker parameters condition the phoneme classification. This approach makes it possible to search for the optimal speaker parameters and the optimal phoneme sequence in a unified optimisation procedure. The general method is tested using speech recognition based on a segmental artificial neural network (ANN). The results from this first test are reported. The performance on the utterance level, measured as the position of the correct phoneme string in the sorted list of highest scoring hypotheses (N-best list) is improved even for very short utterances. The contribution to phoneme classification performance is positive only when recognising long utterances.
Bibliographic reference. Strom, Nikko (1994): "Experiments with a new algorithm for fast speaker adaptation", In ICSLP-1994, 459-462.