Revolutionizing Radiological Analysis: The Future of French Language Automatic Speech Recognition in Healthcare

Area of Science:

• Medical Informatics
• Natural Language Processing
• Radiology

Background:

• Accurate transcription of radiological reports is crucial for patient care and medical documentation.
• Existing Automatic Speech Recognition (ASR) systems often struggle with specialized medical terminology and diverse language accents.
• The need for tailored ASR solutions in specific medical fields and languages is increasingly recognized.

Purpose of the Study:

• To adapt and evaluate the Whisper Large-v2 model for French radiological applications.
• To assess the system's accuracy in transcribing medical terminology and various French accents.
• To demonstrate the potential of specialized ASR to improve efficiency in French-speaking clinical settings.

Main Methods:

• Leveraged the Whisper Large-v2 model for adaptation.
• Focused on French language radiological audio data.
• Involved extensive data collection and preprocessing of medical audio content.
• Evaluated performance using Word Error Rate (WER).

Main Results:

• Achieved a Word Error Rate (WER) of 17.121% for French radiological transcription.
• Demonstrated effective transcription of complex radiological data and diverse accents.
• Validated the system's capability in a specialized medical domain.

Conclusions:

• The adapted Whisper Large-v2 model presents a robust tool for radiological transcription in French.
• This specialized ASR system can significantly enhance medical documentation efficiency.
• Future work includes tailoring ASR for specific medical specialties and languages, and integration with EHRs.