Path Nirvana Sinhala TTS Dataset

High Quality Multi Speaker Sinhala dataset for Text to speech algorithm training - specially designed for deep learning algorithms

Currently there is a lack of publically availble tts datasets for sinhala language of enough length for Sinhala language. This dataset which has 6248 sentences with 13.8 hours of recordings will help mitigate some of these problems. These multi speaker voice recodings were done during the second quarter of 2023. Some effort has been taken to capture most of the rarely used syllables in the Sinhala language, specially those with Sanskrit and Pali origins. There could be some errors (recording not matching with prompt). Though should be rare, if you find some do feel free to contribute to this repository.

There are two speakers

• Male voice Ven. Mettananda - around 5200 voice clips - 11.8 hours in length
• Female voice Mrs. Oshadi - around 1000 voice clips - 2 hours in length

DEMO

I have trained a VITS TTS model using the dataset and samples from the synthesis are now available here

How to use

• You can download the compressed tar file from the Releases which contains the wavs folder with all the audio clips

Stats

• Number of Recordings: 6248
• Total Length: 13.7 hours
• Maximum Length: 15 seconds
• Minumum Length: 2 seconds
• Number of Unique Characters: 54 roman
• List of Roman Characters: !'(),-.:;=?abcdefghijklmnoprstuvyæñāēīōśşūǣḍḥḷṁṅṇṉṛṝṭ
• Silences have been removed from both the beginning and the end of the recordings
• Sample Rate 22050Hz and 16-bit PCM encoded similar to the ljspeech dataset

Output from the create-dataset script for the two releases

v2.0

Total labels => count: 6449, length: 13.9 hours, average length: 7.78
Outliers labels => count: 6399, length: 13.8 hours, average length: 7.78
Used labels => count: 6248, length: 13.7 hours, average length: 7.89
characters=" !'(),-.:;=?abcdefghijklmnoprstuvyæñāēīōśşūǣḍḥḷṁṅṇṉṛṝṭ"
characters=" !'(),-.:;=?[]ංඃඅආඇඈඉඊඋඌඍඑඒඓඔඕඖකඛගඝඞඟචඡජඣඤඥටඨඩඪණඬතථදධනඳපඵබභමඹයරලවශෂසහළෆ්ාැෑිීුූෘෙේෛොෝෞෲ‍‘’“”"

v2.1

Total labels => count: 6449, length: 13.9 hours, average length: 7.78
Outliers labels => count: 6399, length: 13.8 hours, average length: 7.78
Used labels => count: 6386, length: 13.8 hours, average length: 7.77
characters=" !'(),-.:;=?abcdefghijklmnoprstuvyæñāēīōśşūǣḍḥḷṁṅṇṉṛṝṭ"
characters=" !'(),-.:;=?[]ංඃඅආඇඈඉඊඋඌඍඑඒඓඔඕඖකඛගඝඞඟචඡජඣඤඥටඨඩඪණඬතථදධනඳපඵබභමඹයරලවශෂසහළෆ්ාැෑිීුූෘෙේෛොෝෞෲ‍‘’“”"

metadata.csv

• contains a subset of the recordings which are less than 15 seconds long in the ljspeech format

The first version of this dataset contained around 7 hours of speech from a single speaker. It has since been moved to the old dataset folder

See LICENSE.txt file for license information. GPL. The voices belong to a respected buddhist monk and and a lady who donated their voices. Therefore permission is only granted for non-obscene, non-offensive speech generation.

If you have any questions or want to share your TTS training results contact me.

Copyright 2021 Path Nirvana Foundation ([email protected])