Loading datasets...
Release Date: 11/11/2026
Format: MP3, TSV
Size: 119.64 MB
MarmaSpeak TTS is a text-to-speech corpus of read speech in Marma (rmz), a Tibeto-Burman language spoken mainly in the Chittagong Hill Tracts of Bangladesh and in Myanmar, recorded by a single native female speaker. It contains 4,963 recordings (about 5 hours of speech) with their transcripts, recorded in three sessions and split into train (4,615) and test (348). Most prompts were taken from the MarmaSpeak Text Dataset. Audio is MP3 as originally recorded, mostly at 48 kHz. The dataset was created by CLEAR Global with funding from the Australian Government Department of Foreign Affairs and Trade (DFAT). Access is granted on request, to educational institutions and registered non-profit organizations only, with the specific consent of the speaker.
Licensing
Creative Commons Attribution Non Commercial 4.0 International (CC-BY-NC-4.0)
https://spdx.org/licenses/CC-BY-NC-4.0.htmlAccess is granted only on request, and only to requesters with an email address on the domain of an educational institution or a registered non-profit organization; requests from any other email address will not be considered. Eligible requests go through a screening process by CLEAR Global, and every request also requires the specific consent of the speaker, so approval is not guaranteed and may take time. Access is personal to the approved requester and may not be passed on to others.
By downloading the data, the user becomes an independent data controller and is responsible for complying with applicable data protection laws (including GDPR), responding to data subject rights requests, and reporting any data breaches to CLEAR Global. Users must implement appropriate technical and organizational security measures. Data must be retained only as long as necessary for the intended purpose. Users must comply with any data deletion requests from CLEAR Global (including deletion from backups and derived datasets). See the Dataset Access Terms for full details.
Attempting to determine the identity of the speaker in the dataset. Creating synthetic speech that could be used to impersonate the speaker, or for any malicious purposes. Re-hosting or re-sharing the dataset outside the terms of the license.
Ethical Review
The speaker consented to the public release of a sample portion of her recordings only; the full corpus is therefore shared only on request, to educational institutions and registered non-profit organizations, and each request requires her specific consent. The speaker's identity is not disclosed.
Intended Use
Training and evaluation of text-to-speech (speech synthesis) systems for Marma, and research on low-resource speech technology, for non-commercial purposes.
A single native female Marma speaker read sentences in Burmese script, most of them (about 92%) taken from the MarmaSpeak Text Dataset, a collection of Marma sentences compiled by CLEAR Global from textbooks, literature, poems and linguist-authored sentences. The transcripts are the normalized form of each sentence. The recordings were made in three sessions (9, 10 and 11) and are shipped as the original MP3 files.
clips/train/, clips/test/ — 4,963 MP3 files, mono, 32 kbps; 48 kHz except 459 session-9 files at 44.1 kHz.metadata.tsv — one row per clip with the columns path, split, audio_file, session, text, duration, sample_rate.| split | recordings |
|---|---|
| train | 4,615 |
| test | 348 |
| total | 4,963 |
The files total about 9 hours, of which about 5 hours is speech; the rest is leading/trailing silence and pauses.
CLEAR Global trained a TTS model on recording sessions 9 and 10 of this corpus: MarmaSpeak TTS v1, a VITS model fine-tuned from Meta's MMS. It can be tried in the MarmaSpeak TTS demo Space.
@misc{marmaspeak-tts,
author = {CLEAR Global},
title = {MarmaSpeak TTS},
year = {2025}
}
This dataset was created by CLEAR Global as part of a project funded by the Australian Government Department of Foreign Affairs and Trade (DFAT).
Please check README.md for more information.