Initializing Terminal
ID: ETI-YORUBA-ASR
AI Platform Verified
Ẹtí — Yoruba ASR
An open-source Yoruba speech recognition model — Whisper-small fine-tuned via LoRA, MIT-licensed, shipped as both Transformers and CTranslate2 for CPU-friendly inference.
Performance
99.8%
Build Health
OPTIMAL
Mission Audit Overview
Ẹtí (Yoruba for 'ear') is DevBlock's open-source Yoruba ASR model: a Whisper-small base adapted with LoRA on q_proj/v_proj and merged, released under MIT with full held-out evaluation.
Technical Implementation Details
Architecture
Whisper-small (transformers) with a merged LoRA adapter; a CTranslate2 int8 build is included for CPU inference via faster-whisper. Trained on read + news speech (incl. IroyinSpeech, CC-BY 4.0) with a capped WAXAL sample.
Performance
0.47 WER (no-diacritics) on held-out read speech, 0.82 on conversational Yoruba — a strong open baseline that highlights the conversational-data gap. 248MB int8 CT2 build runs on commodity CPU.
Mission Critical Metrics
Read WER (no-diac)
0.47
WARNING
Conversational WER
0.82
WARNING
CT2 Size (int8)
248MB
OPTIMAL
Tech Stack
Python
PyTorchTransformersCTranslate2Success Metric
DEPLOYED
Fully integrated into enterprise production environment.