Automatic Speech Recognition in Primary Progressive Apraxia of Speech

Area of Science:

• Speech-language pathology
• Computational linguistics
• Neurology

Background:

• Diagnosing motor speech disorders like primary progressive apraxia of speech (PPAOS) involves transcribing speech, a labor-intensive process requiring expert listeners.
• Automatic Speech Recognition (ASR) systems offer a potential solution for efficient diagnosis and monitoring of PPAOS.
• Evaluating the efficacy of readily available ASR systems for PPAOS speech transcription is crucial.

Purpose of the Study:

• To assess the effectiveness of the wav2vec 2.0 ASR system in transcribing PPAOS speech.
• To determine if ASR-derived Word Error Rate (WER) can differentiate PPAOS patients from healthy controls and among PPAOS subtypes.
• To investigate the correlation between WER and PPAOS severity, and compare ASR errors to manual transcription errors.

Main Methods:

• Recorded 45 PPAOS patients and 22 healthy controls repeating 13 words.
• Transcribed recordings manually and using the wav2vec 2.0 ASR system.
• Compared WER, phonetic, and prosodic errors between groups and against manual transcriptions.

Main Results:

• Mean WER was significantly higher in PPAOS patients (0.88) than controls (0.33).
• WER correlated with PPAOS severity and distinguished patients from controls, but not PPAOS subtypes.
• ASR phonetic/prosodic errors did not differentiate subtypes, unlike human transcriptions, with poor agreement between ASR and human error counts.

Conclusions:

• ASR, specifically wav2vec 2.0, is valuable for distinguishing disordered from healthy speech and assessing PPAOS severity.
• ASR's current limitations include an inability to differentiate PPAOS subtypes and weak agreement with manual transcriptions.
• ASR can aid PPAOS speech transcription, but its use requires careful consideration of its limitations and the specific research questions.