Universal access: speech recognition for talkers with spastic dysarthria

Harsh Vardhan Sharma, Mark Hasegawa-JohnsonInterspeech 2009 · DOI

Abstract

Automatic Speech Recognition (ASR) offers a promising alternative interface for individuals with motor impairments, yet standard systems often fail to accommodate the acoustic variations of dysarthric speech. This study evaluates speaker-dependent HMM-based recognizers for seven individuals with spastic dysarthria. Results indicate that these customized models can significantly outperform human listeners, particularly for speakers with low intelligibility, by exploiting the predictable nature of articulation errors.

Introduction

For individuals with cerebral palsy and associated spastic dysarthria, traditional input devices like keyboards are often inaccessible. While ASR could provide a "universal access" solution, commercial systems are typically optimized for "typical" speech patterns. Dysarthria results in reduced intelligibility due to muscle weakness and incoordination, causing standard ASR error rates to rise sharply. This research explores whether bespoke, speaker-dependent models can bridge this gap by learning the unique, yet consistent, speech patterns of individual users.

Methods

The researchers implemented three Hidden Markov Model (HMM) configurations to identify the most effective architecture for dysarthric speech:

  1. Whole-word models: Treating each word as a distinct acoustic unit.
  2. Monophone models: Modeling individual phonemes.
  3. Triphone models: Context-dependent phoneme models that account for co-articulation effects.

Acoustic features were represented by 39-dimensional vectors, including 12 Perceptual Linear Prediction (PLP) coefficients, energy, and their respective velocity and acceleration components. The models were optimized by iteratively increasing the number of Gaussian components (up to 32) to maximize accuracy while avoiding over-fitting.

Experimental Setup

The study utilized the UA-Speech database, featuring seven subjects with spastic dysarthria. Each subject recorded 765 isolated words, ranging from simple digits to complex computer commands. To establish a baseline, five naive human listeners transcribed the speech of each subject to determine a "human intelligibility" rating.

Speaker profiles and human intelligibility ratings.

As shown in the table above, the subjects represented a wide spectrum of speech severity, from high intelligibility (86%) to very low (2%).

Results and Analysis

The primary finding of this research is that dysarthric speech, while often unintelligible to humans, is remarkably consistent. ASR systems do not require "correct" pronunciation; they only require that a speaker produces a sound consistently for a given word.

Recognition accuracy across speakers for tasks T07 and T08 using different model types.

Key performance metrics (Percentage of Words Correct, PWC) revealed several insights:

  • Machine vs. Human Performance: For speakers with very low intelligibility, such as F03 (6% human rating), the whole-word ASR system achieved a 40% accuracy on a 55-word vocabulary and 74% on digits. The machine's advantage stems from its speaker-specific training and constrained task vocabulary.
  • Model Suitability: For smaller vocabularies and lower intelligibility speakers, simpler whole-word models often outperformed complex triphone models. However, triphone models demonstrated better scalability as vocabulary size and diversity increased.
  • High Intelligibility Success: For speaker M09 (86% human rating), the system reached nearly 90% accuracy on a 55-word vocabulary, demonstrating that ASR can effectively augment even relatively clear dysarthric speech.

Conclusion and Future Work

This study demonstrates the feasibility of using speaker-dependent ASR as a robust interface for individuals with spastic dysarthria. While the results for isolated words are promising, several challenges remain. Transitioning from isolated words to continuous, connected speech is a significant hurdle. Furthermore, the physical effort required for individuals with dysarthria to provide large amounts of training data necessitates research into more data-efficient adaptation techniques. Future efforts will focus on integrating these models into real-world assistive technologies for text entry and communication.

Summary updated

BibTeX
@inproceedings{sharma2009universal,
  author = {Harsh Vardhan Sharma and Mark Hasegawa-Johnson},
  title = {{Universal access: speech recognition for talkers with spastic dysarthria}},
  booktitle = {Interspeech},
  year = {2009},
  doi = {10.21437/interspeech.2009-444},
}
Figure