Whisper Small — Swahili NSS (AfriVoice 5 Domains + Common Voice)
Stage 2 nonstandard speech fine-tune of openai/whisper-small on the CDLI Kenyan Swahili Nonstandard Speech dataset. Built on top of ElizabethMwangi/whisper-small-swahili-afrivoice-cv, a Stage 1 standard speech checkpoint.
Training Details
- Stage 1 base: ElizabethMwangi/whisper-small-swahili-afrivoice-cv
- Stage 1 data: AfriVoice Swahili all 5 domains (389,599 utterances) + Common Voice Swahili (4,548 utterances)
- Stage 2 dataset: cdli/kenyan_swahili_nonstandard_speech_v1.0 (849 examples)
- Language token: sw
- Full model training: encoder, decoder, and projection updated
- SpecAugment: On
- Learning rate: 5e-5 (polynomial decay, warmup 40 steps)
- Max steps: 1000
- Output dir: sw_nss_tune_whisper_v3_small_afrivoice_v1.2
Results on Kenyan Swahili Nonstandard Speech
Overall
| Metric | Value |
|---|---|
| Overall WER | 0.449 |
| Overall CER | 0.221 |
| Avg WER | 0.365 |
| Avg CER | 0.163 |
| Meaning preservation | 59.4% |
By severity
| Severity | WER | CER |
|---|---|---|
| Mild | 0.29 | 0.12 |
| Moderate | 0.36 | 0.15 |
| Severe | 0.43 | 0.20 |
By etiology
| Etiology | WER | CER |
|---|---|---|
| Cerebral Palsy | 0.44 | 0.21 |
| Multiple Sclerosis | 0.40 | 0.18 |
| Neurodevelopmental disorder | 0.26 | 0.11 |
| Parkinson's Disease | 0.28 | 0.07 |