Luhya ASR Data Subset 50h Whisper Small Finetuned

Luhya-ASR-Data-subset-50h-whisper-small-finetuned

This repository contains a Whisper-family ASR checkpoint trained with the lora_fast profile on a bounded subset of the selected DDD-Kenya dataset.

Model Details

• Owner: Sam4rano
• Base model: openai/whisper-small
• Backend: whisper_finetune
• Dataset: DDD-Kenya/Luhya-ASR-Data-subset-50h
• Output repository: Sam4rano/Luhya-ASR-Data-subset-50h-whisper-small-finetuned
• Sampling rate: 16000 Hz
• Created: 2026-07-12

Intended Use

This model/artifact is intended for automatic speech recognition experiments on Kenyan language speech datasets from DDD-Kenya, especially Gusii and Kamba subset experiments. It is suitable for research evaluation, error analysis, and continued ASR adaptation. It should not be used as the sole basis for high-stakes decisions without additional validation by native speakers and domain experts.

Training and Evaluation Data

The run used bounded, reproducible subsets selected from Hugging Face Datasets. The exact split metadata is saved in run_manifest.json.

Split	Selected examples	Estimated hours
Train	13337	40.00
Validation	400	1.17
Test	1000	2.86

Filters:

{}

Training Procedure

The model was fine-tuned with Hugging Face Seq2SeqTrainer.

• Training profile: lora_fast
• Learning rate: 0.0001
• Per-device train batch size: 2
• Gradient accumulation steps: 8
• Run mode: train
• Epochs: 2.0
• Max steps override: None
• Warmup ratio: 0.05
• FP16 / BF16: False / True
• Gradient checkpointing: False
• Whisper language prompt: None
• Whisper task: transcribe

Results

• Test WER: 46.53%

Additional metrics and raw run metadata are saved alongside the checkpoint:

• test_metrics.json or backend-specific metrics JSON
• run_manifest.json
• prediction CSV files when generated

Limitations

• Luhya is not a native Whisper language-token target; this run avoids forcing a wrong language prompt when WHISPER_LANGUAGE=None.
• The dataset duration metadata is not trusted; selected hours are computed from audio rows that successfully decode.
• Corrupted/unreadable audio rows are skipped and counted in run_manifest.json.
• Subset-based Colab runs are useful for debugging and early experiments, but final claims should use the full validated 50h subset or a fixed documented split.
• Orthography, speaker distribution, audio quality, and dialect coverage should be manually audited before publication.
• Qwen3-ASR baseline rows are comparable evaluation results, not fine-tuned derivative model weights.

Reproducibility

See run_manifest.json for:

• dataset name/config/revision
• base model revision, when resolvable
• split sizes and estimated hours
• Python, PyTorch, CUDA, and GPU metadata
• output directory and repository target

Citation

If you use this artifact, cite the original base model authors and the DDD-Kenya dataset source. For Qwen3-ASR baseline experiments, cite the Qwen3-ASR model card and technical report referenced by the upstream repository.