Iterative LLM-based improvement for French Clinical Interview Transcription and Speaker Diarization

Researchers developed a multi-pass LLM post-processing system that significantly improves French clinical speech transcription accuracy by alternating between speaker recognition and word recognition passes. The system achieved significant word error rate reductions in suicide prevention conversations while maintaining stability in neurosurgery consultations with feasible computational costs for clinical deployment.

• →French medical speech recognition faces challenges with word error rates often exceeding 30% in spontaneous clinical conversations.
• →A multi-pass LLM architecture using Qwen3-Next-80B alternates between speaker recognition and word recognition to improve transcription accuracy.
• →Significant word error rate reductions were achieved in suicide prevention telephone counseling conversations with statistical significance (p < 0.05).
• →The system maintained stability on awake neurosurgery consultations with zero output failures and acceptable computational cost (RTF 0.32).
• →The approach shows feasibility for offline clinical deployment in French healthcare settings.