Whisper Large V3 — Swahili NSS (AfriVoice 5 Domains + Common Voice)
Stage 2 nonstandard speech fine-tune of openai/whisper-large-v3 on the CDLI Kenyan Swahili Nonstandard Speech dataset. Built on top of ElizabethMwangi/whisper-large-v3-swahili-afrivoice-cv, a Stage 1 checkpoint trained on AfriVoice (all 5 domains) interleaved with Common Voice Swahili.
Training Details
- Stage 1 base: ElizabethMwangi/whisper-large-v3-swahili-afrivoice-cv (AfriVoice 5 domains + Common Voice)
- 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
- Best checkpoint: step 550
- Output dir: sw_nss_tune_whisper_v3_large_afrivoice_v1.2
Results on Kenyan Swahili Nonstandard Speech (test set, best checkpoint)
Overall, compared to unadapted baseline
| Model | Overall WER | Avg WER |
|---|---|---|
| openai/whisper-large-v3 (no adaptation) | 0.845 | - |
| Stage 1 only (afrivoice-cv, no NSS tuning) | 0.498 | - |
| This model (adapted) | 0.328 | 0.276 |
| Metric | Value |
|---|---|
| Overall WER | 0.328 |
| Overall CER | 0.158 |
| Avg WER | 0.276 |
| Avg CER | 0.124 |
| Meaning preservation | 76.7% |
By severity
| Severity | WER | CER |
|---|---|---|
| Mild | 0.21 | 0.09 |
| Moderate | 0.25 | 0.11 |
| Severe | 0.34 | 0.15 |