Designing Acoustic Models for Dialectal Bangla: Lessons from 100 Char Clinics

In isolated riparian health clinics across northern Bangladesh, community healthcare providers spend upwards of 40% of their consultation time writing repetitive medical charts by hand.

Standard speech-to-text engines fail drastically in these environments due to background noise from monsoon rains, generator hums, and regional dialectal variations across Pabna, Sirajganj, and Jamalpur.

Field-First Dataset Architecture

Rather than harvesting synthetic studio recordings, our research team embedded with frontline maternal and child healthcare workers to assemble Char-Voice-100:

• Over 350 hours of consent-backed clinical consultations in noise-intensive environments.
• Multi-dialectal phoneme mapping covering Varendra, Rajbanshi, and regional delta inflections.
• Strict on-device, offline-first quantized Conformer architectures designed for low-power ARM tablets.

Accuracy Comparison (Word Error Rate - Lower is better):
- Commercial Cloud Speech API: 42.6% WER
- Fine-tuned Whisper Large-v3: 27.8% WER
- Logicdock Char-Conformer (Edge 4-bit): 11.2% WER

Through this open-source architecture, rural clinics are regaining valuable hours of care every single day.