Combining automatic speech recognition with semantic natural language processing in schizophrenia.

Area of Science:

• Psychiatry
• Computational Linguistics
• Medical Informatics

Background:

• Natural language processing (NLP) tools are vital for quantifying semantic anomalies in schizophrenia.
• Automatic speech recognition (ASR) can accelerate NLP research if sufficiently robust.

Purpose of the Study:

• To assess a state-of-the-art ASR tool's performance.
• To evaluate the impact of ASR on NLP-based diagnostic classification accuracy for schizophrenia.

Main Methods:

• Compared ASR to human transcripts using Word Error Rate (WER) and qualitative error analysis.
• Trained random forest classifiers with semantic similarity measures from both automated and manual transcriptions.
• Evaluated classification accuracy differences between ASR-assisted and manual transcription models.

Main Results:

• The ASR tool exhibited a mean Word Error Rate (WER) of 30.4%, with higher errors for pronouns and sentence-final words.
• Classification accuracy was 76.7% with automated transcriptions versus 79.8% with manual transcriptions.
• The difference in classification performance between automated and manual transcriptions was not statistically significant.

Conclusions:

• Using ASR for semantic analysis in schizophrenia diagnosis results in a minor decrease in accuracy compared to manual transcripts.
• Combining ASR technology with semantic NLP models provides a robust and efficient method for schizophrenia diagnosis.
• ASR facilitates faster NLP research in psychiatric diagnostics without substantial loss of diagnostic power.