Jalak — Indonesian Multi-Speaker TTS Dataset

A Coqui-TTS-ready multi-speaker speech dataset for Indonesian, Javanese, and Sundanese, built to accompany the maiaid/jalak-model VITS checkpoint.

The layout matches jalak-model/config.json exactly: root dataset/ path, Coqui coqui formatter, pipe-separated metadata audio_file|text|speaker_name.

Dataset Summary

Split / metadata file	Speakers	Clips	Source	License
metadata-javanese.csv	39 × JV-xxxxx	5,822	OpenSLR SLR41 (Google + Gadjah Mada Univ.)	CC BY-SA 4.0
metadata-sundanese.csv	41 × SU-xxxxx	4,213	OpenSLR SLR44 (Google + Pendidikan Indonesia Univ.)	CC BY-SA 4.0
metadata-wibowo.csv	1 × wibowo	3,127	cahya/librivox-indonesia, reader 2814 (~4 h)	CC0 / public domain
metadata-ardi.csv	— (placeholder)	0	Microsoft Azure Neural TTS id-ID-ArdiNeural	⚠ ToS-restricted, not redistributed
metadata-gadis.csv	— (placeholder)	0	Microsoft Azure Neural TTS id-ID-GadisNeural	⚠ ToS-restricted, not redistributed
wavs/<speaker>/*.wav	81 dirs	13,162	22 050 Hz mono int16 audio	—

Totals: 13,162 clips · 81 of 83 model speakers · ~12.8 h of audio · 2.4 GB.

The two ardi / gadis speakers are synthetic Azure Neural TTS samples (~2,000 clips each) used in the original training of the jalak model. They are not redistributed here because Azure's Terms of Service prohibit redistribution and commercial use. The pretrained jalak-model already contains both speaker embeddings, so fine-tuning with the 81 speakers in this dataset is sufficient.

Audio

• Sample rate: 22 050 Hz (matches config.json → audio.sample_rate)
• Channels: mono, 16-bit PCM (.wav)
• SLR sources resampled from 48 000 Hz; LibriVox from 44 100 Hz MP3
• Layout: wavs/<speaker_id>/<clip>.wav

Text (IPA phoneme input)

The model consumes IPA phonemes, not plain orthography (config.json → phoneme pipeline run offline). Conversion pipeline:

1. g2p-id (PyPI: g2p-id) — Indonesian grapheme → phoneme (handles schwa ə, glottal stop ʔ, ŋ/ɲ/ʃ, diphthongs aɪ/aʊ/ɔɪ, syllabification via a CRF model).
2. Normalize ASCII g → IPA ɡ (U+0261) to match the model's character vocabulary.
3. Word-initial stress mark ˈ on every token (a deterministic approximation of the original per-syllable stress predictor).
4. Cleanup: q → k, OpenSLR annotation artifacts (p_letter, teng_prep) stripped, é → e.
5. Validation: every character is in the model vocabulary — 0 out-of-vocabulary characters.

Note on fidelity. The original training data applied a per-syllable linguistic stress predictor and an o↔ɔ / e↔ɛ vowel-quality distinction that the g2p-id release does not reproduce. Tokens remain valid for training/fine-tuning; expect minor phonetic simplification relative to the original transcripts.

Structure

jalak/
├── README.md
├── .gitattributes
├── metadata-javanese.csv      # audio_file|text|speaker_name  (5,822 rows)
├── metadata-sundanese.csv     #                               (4,213 rows)
├── metadata-wibowo.csv        #                               (3,127 rows)
├── metadata-ardi.csv          # placeholder (0 rows, documented)
├── metadata-gadis.csv         # placeholder (0 rows, documented)
└── wavs/
    ├── JV-00027/*.wav  ... JV-09724/   # 39 Javanese speakers
    ├── SU-00060/*.wav  ... SU-09757/   # 41 Sundanese speakers
    └── wibowo/*.wav                     # 1 Indonesian audiobook speaker

Each metadata row: wavs/<speaker>/<clip>.wav|<ipa phonemes>|<speaker_id>.

Usage

Quick inference (no training)

Plain text must be converted to IPA first, then fed to the model:

# pip install g2p-id
from g2p_id import G2P
g2p = G2P()
phonemes, _ = g2p.to_phoneme("Selamat pagi, apa kabar?")
# normalize g -> ɡ, then pass `phonemes` as the model's text input

tts --model_path jalak-model/best-model.pth \
    --config_path jalak-model/config.json \
    --speakers_file_path jalak-model/speakers.pth \
    --speaker_idx wibowo \
    --text "<ipa phonemes>" --out_path out.wav

Fine-tuning with Coqui TTS

Point config.json at this dataset root (it already lists all five metadata files under datasets[] with path: "dataset" and the coqui formatter), then run Coqui training.

Reproducibility

The dataset is regenerated from fully public sources by build_dataset.py (parallel resampling + g2p conversion; see phoneme_utils.py). Required downloads:

• https://openslr.org/resources/41/jv_id_female.zip, jv_id_male.zip (SLR41)
• https://openslr.org/resources/44/su_id_female.zip, su_id_male.zip (SLR44)
• https://huggingface.co/datasets/cahya/librivox-indonesia (metadata + audio_train.tgz)

Sources & Citation

• SLR41 / SLR44 — K. Sodimana, K. Pipatsrisawat, L. Ha, M. Jansche, O. Kjartansson, P. De Silva, S. Sarin, "A Step-by-Step Process for Building TTS Voices Using Open Source Data and Framework for Bangla, Javanese, Khmer, Nepali, Sinhala, and Sundanese", Proc. SLTU 2018. http://dx.doi.org/10.21437/SLTU.2018-14
• LibriVox-Indonesia — Cahya / indonesian-nlp (HuggingFace)
• g2p-id — Indonesian Grapheme-to-Phoneme, PyPI
• VITS — J. Kim, J. Kong, J. Son, "Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech", arXiv:2106.06103