Voice ASR Prototype
Clinical Terminology Recognition for Multiple Myeloma
Overview
A voice ASR prototype fine-tuned for clinical terminology recognition, specifically targeting multiple myeloma medical vocabulary that standard ASR models frequently misrecognize.
The Problem
Standard ASR systems struggle with:
- Medical terminology: Drug names, biomarkers, and clinical terms
- Domain-specific vocabulary: Multiple myeloma terminology (BCMA, daratumumab, etc.)
- Clinical workflow integration: On-device deployment requirements
Technical Approach
ASR Fine-tuning
- Fine-tuned OpenAI Whisper Small for on-device deployment
- Trained on recorded patient voice dataset
- Focused on reducing word error rate for myeloma-specific terminology
Domain Vocabulary
- Multiple myeloma drug names (daratumumab, carfilzomib, lenalidomide)
- Biomarker terminology (BCMA, GPRC5D, CD38)
- Clinical assessment terms (VGPR, sCR, MRD negativity)
Results
| Metric | Value |
|---|---|
| Model | Whisper Small |
| Deployment | On-device |
| Focus | Word error rate reduction |
| Domain | Multiple myeloma terminology |
Key Features
- On-device deployment: Optimized for local inference
- Domain-specific training: Patient voice recordings
- Clinical vocabulary: Improved recognition of medical terms
- Standard ASR comparison: Outperforms generic models on clinical terms
Technical Stack
ASR: Fine-tuned Whisper Small
Training Data: Patient voice recordings
Focus: Medical terminology WER
Deployment: On-device
Languages: Python
Applications
- Clinical voice notes
- Medical terminology transcription
- Accessible interfaces for clinical data entry