Beyond manual transcripts: Exploring the potential of automatic speech recognition errors in improving Alzheimer's...

Area of Science:

• Computational linguistics
• Artificial intelligence in healthcare
• Neurodegenerative disease diagnostics

Background:

• Automatic Speech Recognition (ASR) errors are typically viewed as detrimental.
• Previous observations suggest ASR errors may paradoxically aid Alzheimer's Disease (AD) detection.
• A large-scale validation and mechanistic understanding of this phenomenon are needed.

Purpose of the Study:

• To validate the beneficial effect of ASR errors on AD detection.
• To identify the specific mechanisms through which ASR errors aid AD diagnosis.
• To explore the implications for clinical ASR model development.

Main Methods:

• Utilized 18 ASR models (original and fine-tuned) on the ADReSS dataset.
• Generated speech from manual and ASR transcripts using a text-to-speech (TTS) model.
• Employed knowledge-based features and pre-trained embeddings in self-attention and cross-attention AD detection models.
• Conducted detailed analyses of ASR error types, affected words, linguistic features, and attention weights.

Main Results:

• Certain ASR-generated transcripts outperformed manual transcripts in AD detection accuracy.
• Performance gains were linked to specific, asymmetric ASR error patterns, not high Word Error Rate (WER).
• These error patterns amplified existing linguistic deficits in AD speech, increasing group divergence.
• Diagnostic clues were preserved in synthesized speech, supporting data augmentation potential.

Conclusions:

• Specific ASR error patterns enhance AD detection by amplifying pathological linguistic deficits.
• This suggests a shift in ASR development towards diagnostic utility over pure transcription accuracy.
• Findings have implications for developing ASR tools for neurodegenerative disease research and clinical applications.