Whisper Small — Swahili (Non-Standard Speech, No SpecAugment)
Fine-tuned version of ElizabethMwangi/whisper-small-swahili-afrivoice_cv for automatic speech recognition of non-standard (dysarthric) Swahili speech.
This is Stage 2 of a two-stage fine-tuning pipeline, further trained on the Kenyan Swahili Non-Standard Speech dataset.
Model Details
- Base model: ElizabethMwangi/whisper-small-swahili-afrivoice_cv
- Language: Swahili (
sw) - Task: Automatic Speech Recognition (ASR)
- Training data: Kenyan Swahili Non-Standard Speech v1.0
Evaluation Results (NSS test set, 849 examples)
| Metric | Value |
|---|---|
| Overall WER (normalised) | 45.5% |
| Overall CER (normalised) | 21.7% |
| Avg WER (normalised) | 36.8% |
| Avg CER (normalised) | 16.2% |
Per-severity breakdown
| Severity | Avg WER | Avg CER |
|---|---|---|
| Mild | 0.29 | 0.12 |
| Moderate | 0.36 | 0.15 |
| Severe | 0.44 | 0.20 |
Usage
from transformers import pipeline
asr = pipeline("automatic-speech-recognition", model="ElizabethMwangi/whisper-small-swahili-nss_afrivoice_cv")
result = asr("audio.wav")
print(result["text"])
Training Details
- Fine-tuned using Hugging Face
Seq2SeqTrainer - Language token:
sw - Learning rate: 5e-5 with polynomial decay
- Batch size: 16
- Total steps: 1000
- Spec augment: disabled (no SpecAugment)
Citation
If you use this model, please cite the AfriVoice and Kenyan Swahili Non-Standard Speech datasets and this repository.