Yohan2003/whisper-sl-data

Whisper Sinhala Speech Dataset

Sinhala speech-to-text data aggregated from multiple sources for fine-tuning Whisper-family ASR models. The dataset combines audio from OpenSLR-52, YouTube, BizBrains, and Linga sources, and is provided as four different train/test/validation splits of the same underlying pool of examples.

Trained models using this data: Yohan2003/whisper-small-sinhala (also bundles a copy of these same splits).

Dataset Structure

Each split is a folder containing train.parquet, test.parquet, and validation.parquet, with the following columns:

Column	Type	Description
audio	binary	Raw audio bytes
text	string	Transcription
source_dataset	string	Origin of the sample: openslr, youtube, bizbrains, or linga

Splits

All splits are drawn from the same ~154,828-example pool but partitioned differently:

stratified/ (v1)

Stratified split by source_dataset proportions.

• train: 123,862
• test: 15,483
• validation: 15,483

stratified_v2/

Regenerated stratified split (same methodology as v1, refreshed data/processing).

• train: 123,862
• test: 15,483
• validation: 15,483

stratified_v3/

Speaker-disjoint split — speakers are constrained to appear in only one of train/test/validation, preventing speaker leakage across splits.

• train: 123,205
• test: 15,860
• validation: 15,763

data/stratified_v4/

Same row membership as stratified_v3 (same speaker-disjoint train/test/validation assignment; audio and source_dataset are unchanged), with a spacing normalization pass applied to text: a 1,046-rule map that fixes Sinhala compound words written inconsistently as either one joined token or two spaced tokens (e.g. කළහැකි vs කළ හැකි) across the whole corpus, so the same word is spelled the same way everywhere. Rows changed: train 4,148 / validation 525 / test 495.

• train: 123,205
• test: 15,860
• validation: 15,763
• Note: this split lives under data/ (data/stratified_v4/...), unlike v1-v3 which are at the repo root.

Source composition (train, consistent across splits): ~97% OpenSLR-52, ~2% YouTube, ~0.6% BizBrains, ~0.6% Linga.

Intended Use

Fine-tuning and evaluating Sinhala ASR models (e.g. Whisper). data/stratified_v4/ is the recommended default: it has the same speaker-disjoint guarantee as stratified_v3 plus spacing-normalized text, which removes a class of spurious WER errors caused by the reference transcripts themselves being inconsistently spaced (the same word written two different ways). Use stratified_v3 if you specifically need the un-normalized text for comparison. stratified/stratified_v2 (v1/v2) may share speakers across train and eval splits and are kept for backward compatibility only.

Source Data

• OpenSLR-52: Sinhala speech corpus (https://openslr.org/52/)
• YouTube: Scraped Sinhala speech with transcripts
• BizBrains, Linga: Additional Sinhala speech sources