State-Transition Interpolation and MAP Adaptation for HMM-based Dysarthric Speech Recognition
Harsh Vardhan Sharma, Mark Hasegawa‐Johnson · 2010
This paper describes the results of our experiments in building speaker-adaptive recognizers for talkers with spastic dysarthria. We study two modifications – (a) MAP adaptation of speaker-independent systems trained on normal speech and, (b) using a transition probability matrix that is a linear interpolation between fully ergodic and (exclusively) leftto-right structures, for both speaker-dependent and speaker-adapted systems. The experiments indicate that (1) for speaker-dependent systems, left-to-right HMMs have lower word error rate than transition-interpolated HMMs, (2) adapting all parameters other than transition probabilities results in the highest recognition accuracy compared to adapting any subset of these parameters or adapting all parameters including transition probabilities, (3) performing both transition-interpolation and adaptation gives higher word error rate than performing adaptation alone and, (4) dysarthria severity is not a sufficient indicator of the relative performance of speakerdependent and speaker-adapted systems. Automatic speech recognition (ASR) systems generally assume that the speech signal is a realisation of some message encoded as a sequence of one or more symbols. To effect the reverse operation of recognising the underlying symbol sequence given a spoken utterance, the continuous speech waveform is first converted to a sequence of equally spaced discrete parameter vectors. The role of the recogniser is to effect a mapping between sequences of speech vectors and the wanted underlying symbol sequences. Most speech recognizers today are based on the hidden Markov model (HMM) paradigm: it is assumed that the sequence of observed speech vectors is generated by a Markov model as shown in Fig. 1. A Markov model is a finite state machine which changes state once every time unit and each time t that a state j is entered, a speech vector ot is generated from the probability density bj(ot) which