Prompt Setup
- Prompt mode:
metadata
- Method note: This is a metadata-conditioned/oracle prompt experiment. The prompt uses speaker severity, disorder type, and etiology from
speaker_metadata.csv; compare it separately from deployment baselines.
- WER framing: this notebook does not ask the model to remove repetitions, clean disfluencies, summarize, or normalize grammar.
Test Metrics
- Normalized WER: 0.632689
- Normalized CER: 0.340070
- Avg capped WER: 0.491120
- Avg capped CER: 0.229225
Training Setup
- Dataset:
cdli/ugandan_luganda_nonstandard_speech_v1.0
- Splits: cleaned TSVs matched by
audio_id
- Base model:
KasuleTrevor/cdli-qwen3-asr-lg-typical-1p7b-base-finetune
- LR:
0.0001
- Scheduler:
constant_with_warmup
- Epochs:
3
- Train mode:
full
- Seed:
42
Speaker speech disorder type, speech impairment severity, and etiology are inserted from metadata. The prompt then instructs exact Luganda transcription while preserving repetitions, disfluencies, false starts, and partial words.
