Whisper Large V3 - Swahili Afrivoice (All 5 Domains)
Fine-tuned on all 5 domains of Swahili Afrivoice dataset (389,599 utterances).
Training Details
| Parameter | Value |
|---|---|
| Base Model | openai/whisper-large-v3 |
| Language | Swahili (sw) |
| Task | Transcribe |
| Dataset | All 5 domains (389,599 utterances) |
| Max Steps | 1000 |
| Best Checkpoint | Step 600 |
| Batch Size | 4 (effective batch 32) |
| Learning Rate | 1e-5 |
| SpecAugment | Enabled |
Results
| Dataset | WER |
|---|---|
| Afrivoice (Avg WER) | 10.9% |
| Common Voice Swahili | 29.7% |
| Non-standard Speech | 43.4% |
Usage
from transformers import pipeline
asr = pipeline("automatic-speech-recognition", model="ElizabethMwangi/whisper-large-v3-swahili-afrivoice-5domains")
result = asr("audio.wav")
print(result["text"])