SinhalaTTS_4speaker

Multi-speaker Sinhala text-to-speech corpus (4 native speakers), prepared for training multi-speaker TTS models such as VITS.

Structure

• Splits: train (80%) / test (20%), stratified per speaker (seed 42).
• Audio: mono, 16-bit PCM WAV, 22050 Hz (resampled from 44.1/48 kHz originals).

Column	Description
audio	Speech clip (playable in the viewer)
text	Transcript as provided by the source. Sinhala Unicode script, except speaker Isuru, whose source transcripts are romanized only
text_romanized	Romanized (ISO 15919-style) transcript. Empty for Harini (not provided by the source)
speaker_id	Integer speaker id 0–3 (ClassLabel)
speaker_name	Dinithi, Harini, Isuru, Yasindu
gender	Speaker gender
duration	Clip length in seconds
id	Original utterance id

Statistics

Split	Speaker	ID	Gender	Utterances	Hours
train	Dinithi	0	female	2000	3.85
train	Harini	1	female	915	1.71
train	Isuru	2	male	460	0.98
train	Yasindu	3	male	1040	1.73
test	Dinithi	0	female	500	0.98
test	Harini	1	female	229	0.43
test	Isuru	2	male	115	0.24
test	Yasindu	3	male	260	0.44

Usage

from datasets import load_dataset

ds = load_dataset("TriLingDATA/SinhalaTTS_4speaker")
sample = ds["train"][0]
print(sample["text"], sample["speaker_id"], sample["audio"]["sampling_rate"])

For VITS multi-speaker training use speaker_id as the speaker index (n_speakers=4). Since Isuru's transcripts are romanized while the other speakers use Sinhala script, either train on text_romanized where available, transliterate consistently, or filter by speaker so the model sees a single script.

Notes

• 56 Harini recordings in the source release are corrupt (unreadable WAV headers) and were excluded.
• Duplicate metadata rows in Isuru's source were de-duplicated.

Source & credit

Original recordings and transcriptions by Safnas Kaldeen (Final Year Project, Department of Computer Science and Engineering, University of Moratuwa): https://github.com/SafnasKaldeen/TTSx. Please cite the original repository when using this data.