Whisper Small — Swahili (AfriVoice + CommonVoice, No SpecAugment)
Fine-tuned version of openai/whisper-small for automatic speech recognition of Swahili, trained on standard speech from the AfriVoice Swahili dataset and CommonVoice Swahili.
This is Stage 1 of a two-stage fine-tuning pipeline. Stage 2 further fine-tunes this model on non-standard Swahili speech — see ElizabethMwangi/whisper-small-swahili-nss_afrivoice_cv.
Model Details
- Base model: openai/whisper-small
- Language: Swahili (
sw) - Task: Automatic Speech Recognition (ASR)
- Training data: AfriVoice Swahili + CommonVoice Swahili
Evaluation Results (AfriVoice Swahili test set)
| Metric | Value |
|---|---|
| WER (adjusted, normalised) | 47.41% |
| CER (adjusted, normalised) | 19.82% |
| WER (unadjusted, normalised) | 48.02% |
| CER (unadjusted, normalised) | 19.99% |
Usage
from transformers import pipeline
asr = pipeline("automatic-speech-recognition", model="ElizabethMwangi/whisper-small-swahili-afrivoice_cv")
result = asr("audio.wav")
print(result["text"])
Training Details
- Fine-tuned using Hugging Face
Seq2SeqTrainer - Language token:
sw - Learning rate: 1e-5 with polynomial decay
- Batch size: 32
- Total steps: 1000
- Spec augment: disabled (no SpecAugment)
Citation
If you use this model, please cite the AfriVoice Swahili dataset and this repository.