01The problem
Yoruba transcription can lose meaningful tone marks even when the underlying letters look plausible. A single word-error score can hide that distinction.
02The approach
- Audit the WAXAL Yoruba yor_tts subset, remove held-out transcript duplicates from training and retain the published full test set.
- Fine-tune Whisper tiny, base and small for 2,000 steps each. Choose checkpoints using validation WER, never test results.
- Convert selected checkpoints to int8, q5_0 and q4_0 and compare with runtime-specific floating-point controls. Report tone errors together with lexical and tone coverage.
03The evidence
- All three training runs and their 177-recording GPU evaluations completed.
- The published test split contains 1.562 measured audio hours.
- The MMS baseline achieved 58.41% tone-stripped WER. Validation-selected fine-tuned tiny achieved 88.76%; lower is better.
04Scope & status
Compression scoring and controlled serving-cost analysis are unfinished. The speech was collected for TTS, speakers overlap across splits, and current recognition accuracy is weak. These are preliminary within-corpus results, not a production ASR endorsement.
Public source is available for independent inspection.