Abstract
Automatic Speech Recognition (ASR) for individuals with dysarthria faces significant challenges due to high speech variability and limited training data. This paper investigates the efficacy of Maximum A Posteriori (MAP) adaptation and Transition-Interpolation (TI) within a Hidden Markov Model (HMM) framework. By evaluating 16 system configurations on the UA-Speech database, the study finds that while TI aims to capture irregular speech patterns, standard left-to-right models remain superior. The most effective approach involves MAP adaptation of acoustic parameters while preserving the transition probabilities of a "normal" speech model.
Introduction
Dysarthria is a motor speech disorder resulting from neurological damage (such as cerebral palsy) that impairs the muscles used for speaking. For these individuals, ASR is a vital assistive technology, yet standard systems struggle with the highly variable nature of dysarthric speech, which often features unpredictable pauses, repetitions, and "outlier events." Furthermore, collecting sufficient data for speaker-dependent models is difficult because prolonged speaking is physically exhausting for these users. This research explores building accurate recognizers with limited data by comparing speaker-dependent (SD) and speaker-adapted (SA) systems.
Methodology
The authors investigated two primary modifications to the standard HMM framework to better accommodate dysarthric speech patterns:
MAP Adaptation
Instead of training models from scratch, the researchers utilized a "MAP prior" trained on the TIMIT corpus (standard, non-disordered speech). This prior was then adapted using the dysarthric speaker's data. They tested 16 configurations (C00–C15) adapting different combinations of Gaussian means ($m$), variances ($v$), mixture weights ($w$), and transition probabilities ($t$).
Transition-Interpolation (TI)
Standard ASR models typically follow a "strictly left-to-right" (LR) topology, assuming speech moves in a linear sequence. Transition-Interpolation (TI) models are a hybrid between LR and ergodic models, allowing the system to theoretically capture non-linear progress, such as a speaker struggling through a syllable or pausing mid-word.
Results and Key Insights
The "Rigidity" Benefit
Surprisingly, transition interpolation generally performed worse than standard left-to-right models. While TI was designed to capture outlier events, it likely introduced too much "slack" or uncertainty into the model, leading to more errors than it solved. In SD systems, the LR model (C00) outperformed the TI model (C01) for six out of seven speakers.
Optimal Adaptation Strategy
The most successful configuration, C13, involved adapting acoustic parameters (means, variances, and weights) while leaving the transition probabilities from the "normal" speech model untouched. This suggests that the underlying temporal structure of phones remains somewhat consistent, even if the sounds themselves are distorted.
Performance Across Severity Levels
Speakers with very low intelligibility saw the most significant gains from speaker adaptation. For these users, a speaker-dependent model often fails because their speech is too inconsistent to provide a clean training signal; the "normal" speech prior provides a necessary structural backbone. Notably, for speakers with very low intelligibility (e.g., M06), the best ASR systems actually performed better than the naive human listeners in the study.
Conclusion
This research demonstrates that for spastic dysarthria, maintaining a rigid temporal structure while adapting acoustic parameters is the most effective modeling strategy. While ASR still lags behind human listeners for mid-to-high intelligibility speakers, these adaptations offer a significant path forward for those with the most severe speech impairments. Future work will focus on extending these techniques to continuous, conversational speech.
Summary updated