Parameter	Value
Base Model	openai/whisper-large-v3
Language	Swahili (sw)
Task	Transcribe
Dataset	Afrivoice Swahili - Education Domain
Max Steps	1000
Max Epochs	5
Batch Size	4
Gradient Accumulation	8 (effective batch 32)
Learning Rate	1e-5
LR Scheduler	polynomial (power: 4)
LR Warmup Steps	100
LR End	1e-8
FP16	True
SpecAugment	Enabled
Save Steps	50
Evaluation Steps	50
Max Generation Length	128

Results

Dataset	WER	CER
Afrivoice Swahili (Dev)	13.5%	4.2%
Afrivoice Swahili (Test)	13.5%	4.2%

Zero-shot on Non-Standard Speech

Dataset	WER	CER
Kenyan Swahili Non-Standard Speech	46.7%	21.7%

Usage

from transformers import pipeline

asr = pipeline("automatic-speech-recognition", model="ElizabethMwangi/whisper-large-v3-swahili-afrivoice") result = asr("audio.wav") print(result["text"])

Training Checkpoints

Saved to: /jupyter_kernel/trained_models/sw_afrivoice_tune_whisper_large_2/best_model

Citation

@misc{afrivoice_swahili_whisper, title={Whisper Large V3 Fine-tuned on Swahili Afrivoice Education Domain}, author={Elizabeth Mwangi}, year={2026} }