Whisper Small — Swahili (AfriVoice 5 Domains + Common Voice)
Stage 1 standard speech fine-tune of openai/whisper-small on AfriVoice Swahili (all 5 domains) interleaved with Common Voice Swahili. This model is trained on standard speech and serves as the base checkpoint for Stage 2 nonstandard speech fine-tuning (ElizabethMwangi/whisper-small-swahili-nss-afrivoice-cv).
Training Details
- Dataset: AfriVoice Swahili all 5 domains (389,599 utterances) interleaved with Common Voice Swahili (4,548 utterances)
- Language token: sw
- Full model training: encoder, decoder, and projection updated
- SpecAugment: On
- Learning rate: 1e-5 (polynomial decay, warmup 100 steps)
- Max steps: 1000
- Best checkpoint: step 650 (lowest eval WER 0.481)
- Output dir: sw_afrivoice_tune_whisper_small_v0.2
Results on AfriVoice Swahili (Standard Speech)
| Split | WER | CER |
|---|---|---|
| Dev | 0.481 | 0.197 |
| Test | 0.475 | 0.199 |
Results on Kenyan Swahili Nonstandard Speech (unadapted, test set)
These results show performance on nonstandard speech before Stage 2 fine-tuning.
| Metric | Value |
|---|---|
| Overall WER | 0.596 |
| Overall CER | 0.248 |