SunflowerASR-Africa51-GeoPoll

This model was built specifically to improve Swahili speech recognition on telephone-quality audio. It builds on Sunbird/SunflowerASR-51-african-languages and was fine-tuned on GeoPoll telephone survey calls for better accuracy.

Training data

GeoPoll telephone survey recordings in Swahili with human transcriptions. The model trained on 11,745 training clips, about 27 hours.

Training

Setting	Value
Method	LoRA
LoRA rank	32
Encoder trained	yes
Learning rate	5e-05
Warmup steps	500
Steps	6000
Batch size	4
Gradient accumulation	4
Augmentation probability	0.5
Shortest clip (s)	2
Seed	42

Evaluation

Word and character error rates before fine-tuning (Sunbird/SunflowerASR-51-african-languages) and after (this model), scored with GeoPoll's text normalization. Lower is better.

Test set	Language	Clips	WER before	WER after	CER before	CER after	WER reduction
GeoPoll held-out v4 (sw)	Swahili	5,214	31.7%	15.1%	13.2%	6.8%	53%

GeoPoll held-out v4 (sw): 5,214 segments (10.3 hours) of real survey calls with 40 respondents in Kenya and Tanzania. These calls were held out: none of their audio was used to train the model. Every segment is at least 2 seconds long.

How to use

from transformers import pipeline

asr = pipeline("automatic-speech-recognition", model="GeoPoll/SunflowerASR-Africa51-GeoPoll", chunk_length_s=30)
options = {"language": "sw", "task": "transcribe"}
text = asr("call.wav", generate_kwargs=options)["text"]