Modelpublic

cdli/whisper-large-v3_finetuned_kinyarwanda_standard_speech_v0.9

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

0likes46downloads

Model Card

Whisper Large V3 - Kinyarwanda (Standard Speech)

Fine-tuned version of openai/whisper-large-v3 for automatic speech recognition on standard Kinyarwanda speech.

Note on language token

Whisper's supported language set does not include Kinyarwanda. This model was trained using the sw (Swahili) language token as a substitute, since it is one of the closer supported languages available. Use `language="sw"` at inference time — this is expected and required, not an error.

Dataset

Source data is DigitalUmuganda/Afrivoice_Kinyarwanda (CC-BY-4.0), a Kinyarwanda ASR dataset from Digital Umuganda. The training dataset used here, ElizabethMwangi/kinyarwanda_afrivoice_all_domains_v0.1, was created by extracting and concatenating 5 of its domains:

—Agriculture
—Education
—Financial
—Government
—Health

Note: scripted_education subset was excluded due to corrupted files.

Training Configuration

Setting	Value
Base model	openai/whisper-large-v3
Language token used	sw (Kinyarwanda is not natively supported by Whisper; see note above)
Task	transcribe
Encoder updated	Yes
Decoder updated	Yes
Projection updated	Yes
Trainable parameters	1,543,490,560 (full model)
Batch size	32
Learning rate	1e-5
LR scheduler	polynomial (power=4, warmup=100 steps, end=1e-8)
Precision	fp16
Max steps	500

Results

Evaluated on held-out validation data at step 500:

Metric	Value
Validation Loss	0.2997
WER	0.2044
CER	0.0626

WER and CER improved consistently from step 0 (WER 0.93, CER 0.34) through step 500, with both training and validation loss decreasing steadily and no signs of overfitting at this point in training.

Usage

python

from transformers import WhisperProcessor, WhisperForConditionalGeneration

processor = WhisperProcessor.from_pretrained("ElizabethMwangi/whisper-large-v3_finetuned_kinyarwanda_standard_speech_v1.0", language="sw", task="transcribe")
model = WhisperForConditionalGeneration.from_pretrained("ElizabethMwangi/whisper-large-v3_finetuned_kinyarwanda_standard_speech_v1.0")

Limitations

This model was trained on standard (non-impaired) Kinyarwanda speech across five domains. Performance on non-standard or impaired speech, or domains outside those listed above, has not been evaluated. Because Whisper does not natively support Kinyarwanda, this model repurposes the sw language token