SinhaSpeech/sinhala-asr-data

SinhaSpeech Sinhala ASR Data

Sinhala speech-to-text data aggregated from multiple sources for fine-tuning Whisper-family ASR models. It combines audio from OpenSLR-52, YouTube, BizBrains, and Linga, and is provided as seven versioned train/validation/test splits (stratified_v1 to stratified_v7) of the same underlying pool of ~154,828 examples.

Models trained on this data: SinhaSpeech/whisper-small-sinhala-v6-e6-run11-best (best, 16.38% test WER), SinhaSpeech/whisper-small-sinhala-v6-e5, and the earlier runs in SinhaSpeech/whisper-small-sinhala-experiments.

Dataset Structure

Every split lives at data/stratified_vN/ and contains train.parquet, validation.parquet and test.parquet:

Column	Type	Description
audio	binary	Raw audio bytes
text	string	Transcription
source_dataset	string	Origin of the sample: openslr, youtube, bizbrains, or linga
language	string	Language of the sample: si (Sinhala) or en (English). Present in stratified_v7 train only.

Splits

Split	Train	Validation	Test	Speaker-disjoint	What changed
data/stratified_v1	123,862	15,483	15,483	No	Stratified by source_dataset proportions
data/stratified_v2	123,862	15,483	15,483	No	Regenerated v1 (same method, refreshed processing)
data/stratified_v3	123,205	15,763	15,860	Yes	Speakers appear in only one of train/validation/test
data/stratified_v4	123,205	15,763	15,860	Yes	v3 membership plus spacing normalization of text (1,046-rule map for compounds written inconsistently as joined vs spaced, e.g. කළහැකි / කළ හැකි; rows changed: train 4,148 / val 525 / test 495)
data/stratified_v5	123,205	15,763	15,860	Yes	v4 plus register and word-boundary normalization of text
data/stratified_v6	123,205	15,763	15,860	Yes	v5 plus canonicalization of ~2,954 word instances: letter-level spellings (ළ/ල, ණ/න), ZWJ/rakaransaya conjuncts, word-level register pairs (e.g. -නං/-නම්), and English typos in code-switched references
data/stratified_v7	108,278	14,084	15,860	Yes	v6 plus text-disjoint constraint (no sentence shared across splits) and 10,000 English (LibriSpeech train-clean-100) rows mixed into train for anti-forgetting. Adds language column (si/en).

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

WER is only comparable between models evaluated on the same test split. v1/v2 share a test set, and v3 to v6 share the same test rows with progressively cleaner references.

Intended Use

Fine-tuning and evaluating Sinhala ASR models. data/stratified_v7 is the recommended default. v1/v2 may share speakers across train and evaluation splits and are kept for reproducibility of the earliest runs only. Use v3 if you need the un-normalized text.

from huggingface_hub import hf_hub_download
p = hf_hub_download("SinhaSpeech/sinhala-asr-data", "data/stratified_v6/test.parquet", repo_type="dataset")

Source Data

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