teckedd/gha-whisper-small-twi-v6

Twi/Akan + English retention Whisper ASR for Ghana Health AI (v6-local-holdout). Recipe v6 mix; promote=False.

• Base model: teckedd/gha-whisper-small-twi-v6
• Task: automatic-speech-recognition
• Languages: tw, ak, en
• Card generated: 2026-08-16 (UTC)
• Product: ghanahealth.serendepify.com

Not a medical device. Outputs support community health guidance only.

Metrics

Metric	Value
val_wer	0.2970
val_cer	0.1031

Training data

• google/WaxalNLP (aka_asr)
• local (ghana-health-ai-recorder)
• fsicoli/common_voice_22_0 (tw)
• fsicoli/common_voice_22_0 (en)
• ghananlpcommunity/twi-speech-text-multispeaker-16k (default)

Intended use

• In-product ASR / TTS / chat for Twi (Akan) and English health conversations in Ghana.
• Research on low-resource Ghanaian language speech and health dialogue.

Out of scope

• Clinical diagnosis or autonomous medical decisions.
• Claiming near-native quality without reporting held-out WER/CER or human A/B scores.

How to load

from transformers import pipeline  # or AutoModel + processor per task
# repo: teckedd/gha-whisper-small-twi-v6

Citation

If you use this checkpoint, please credit Ghana Health AI / Serendepify and the upstream base model authors plus any listed datasets.

Recipe

• Run: v6-local-holdout
• Freeze encoder: False
• LR / steps / batch: 5e-06 / 600 / 8x4
• Local recorder data: True · manifest /root/gha_local_asr/manifest.train32.jsonl · weight 0.08
• Baseline to beat (v6 serving, Waxal full test greedy): 0.3149
• Beats v6 greedy: False · v6 beam5 bar (0.3044): False
• Historical reference (Round 2 greedy, superseded): 0.3283

Serving

Production decode uses num_beams=5 in modal/asr_service.py.