Whisper Large V3 — Swahili (AfriVoice All Domains + Common Voice)
Fine-tuned version of openai/whisper-large-v3 for Swahili ASR, trained on all 5 AfriVoice domains combined with CommonVoice Swahili.
Training Details
- Base model: openai/whisper-large-v3
- Language: Swahili (
sw) - Task: Automatic Speech Recognition (ASR)
- Training data: ElizabethMwangi/swahili_afrivoice_all_domains_v0.1 (389,599 utts) + cdli/common_voice_swahili_small (4,548 utts) = 394,147 total
- Full model training: encoder, decoder, and projection updated
- SpecAugment: On
- Learning rate: 1e-5 (polynomial decay, LR end 1e-8, power 4)
- Effective batch size: 32 (batch 4 x grad accum 8)
- Max steps: 1000 (run in two stages: 0-400, then 400-1000)
- Best checkpoint: Step 700 (AfriVoice dev WER 0.2393)
- Output dir: sw_afrivoice_cv_tune_whisper_large_v0.1
Results
| Dataset | WER | CER |
|---|---|---|
| AfriVoice dev | 0.239 | 0.089 |
| AfriVoice test | 0.244 | 0.094 |
| CommonVoice Swahili test | 0.201 | 0.049 |
| Swahili NSS test (885 examples) | 0.394 | 0.172 |
Usage
from transformers import pipeline
asr = pipeline(
'automatic-speech-recognition',
model='ElizabethMwangi/whisper-large-v3-swahili-afrivoice-cv'
)
result = asr('audio.wav')
print(result['text'])
Intended Use
Swahili ASR with a focus on Kenyan speakers, including nonstandard speech patterns. Part of CDLI research into inclusive ASR for African languages.
Citation
If you use this model, please cite the AfriVoice dataset (DigitalUmuganda/Afrivoice_Swahili) and this repository.