Whisper Small Sinhala Prototype

This repository contains an experimental fine-tuned checkpoint of OpenAI Whisper Small for Sinhala automatic speech recognition. It was trained with the large-sinhala-asr-dataset and is published as a prototype for evaluation and continued development.

Model details

Intended use

Use this checkpoint for research, baseline comparison, testing, and further fine-tuning. It should not be treated as a validated production model until dataset documentation and held-out evaluation results are added.

Quick start

import torch
from transformers import pipeline

model_id = "kaan84/whisper-small-sinhala-proto"
device = 0 if torch.cuda.is_available() else -1

transcriber = pipeline(
    task="automatic-speech-recognition",
    model=model_id,
    device=device,
)

result = transcriber(
    "sample.wav",
    generate_kwargs={"language": "si", "task": "transcribe"},
)
print(result["text"])

Training data

The model metadata identifies large-sinhala-asr-dataset as the training dataset. For reproducibility, document its repository link, source, license, total duration, speaker distribution, split strategy, audio sampling rate, transcript normalization, and quality-control process.

Training procedure

The checkpoint was generated using the Transformers Trainer workflow. Add the exact learning rate, batch size, gradient-accumulation steps, epoch or step count, warm-up schedule, precision, hardware, seed, and checkpoint-selection rule.

Evaluation

This prototype was fine-tuned using 3,000 Sinhala speech samples and prepared with an 800-sample test subset.

A qualitative test example produced a sample-level WER of 50.00% and CER of 11.11%. These values are based on one notebook example and must not be interpreted as corpus-level performance.

Full evaluation across the 800-sample test subset was not completed in the available experiment. This model should therefore be treated as a research prototype rather than a production-ready ASR system.

Limitations and responsible use

• This is an experimental checkpoint and may generate unreliable transcripts.
• Results can degrade with background noise, code-switching, overlapping voices, or unfamiliar accents.
• Training-data coverage may not represent all Sinhala-speaking communities or domains.
• Review outputs manually before using them for consequential decisions or official records.

License

Apache License 2.0. Users must also comply with the licenses and terms of the base model and training data.

Version relationship

This repository should remain clearly labeled as a prototype. If kaan84/whisper-small-sinhala-v1 is the selected release model, direct general users to that repository and retain this checkpoint for comparison and reproducibility.