Chalermdej/yodas2_sidon_th_tts

Thai TTS Dataset — Filtered & Quality-Verified from YODAS2 sidon

A filtered, quality-verified Thai text-to-speech dataset derived from sarulab-speech/yodas2_sidon, with transcriptions verified by multiple ASR models and Gemini, text fully normalized to Thai, and audio quality-screened with DNSMOS.

Dataset Summary

Samples	141,927
Audio hours	156.0
Speakers	4,199
Sample rate	24,000 Hz
Format	WAV, PCM 16-bit, mono
Language	Thai
Source	sarulab-speech/yodas2_sidon
Download size	~26 GB
License	CC-BY-3.0

Usage

from datasets import load_dataset

ds = load_dataset("Chalermdej/yodas2_sidon_th_tts")

# Stream without downloading everything
ds = load_dataset("Chalermdej/yodas2_sidon_th_tts", streaming=True)

sample = ds["train"][0]
print(sample["text"])        # Thai text (fully normalized)
print(sample["speaker_id"])  # Speaker identifier (YouTube video ID)
print(sample["grade_avg"])   # Quality grade string: "S+", "S", or "A+"
print(sample["audio"])       # {"array": [...], "sampling_rate": 24000}

Filtering by quality grade

grade_avg is a string ("S+", "S", "A+"), not a number. Filter like this:

from datasets import load_dataset

ds = load_dataset("Chalermdej/yodas2_sidon_th_tts")

# High quality only (S+ and S — ~130k samples, ~143 hours)
ds_clean = ds.filter(lambda x: x["grade_avg"] in ["S+", "S"])

# Highest quality only (S+ — ~78k samples, ~82 hours)
ds_best = ds.filter(lambda x: x["grade_avg"] == "S+")

Dataset Structure

Column	Type	Description
audio	Audio	24kHz mono WAV
utt_id	string	Unique utterance ID (source video ID + timestamps)
text	string	Normalized Thai text (no English words, no raw numbers)
speaker_id	string	YouTube video ID — one video ≈ one speaker; consistent across the dataset
grade_avg	string	Quality grade: "S+", "S", or "A+" — always a string, not a number
duration	float	Duration in seconds
dnsmos_overall	float	DNSMOS overall score
dnsmos_signal	float	DNSMOS signal score
dnsmos_background	float	DNSMOS background score

Quality Grades

Grades are based on average transcription agreement across 3 ASR models:

Grade	ASR Agreement	Samples	Hours
S+	95–100%	78,180	82.4h
S	90–95%	51,518	60.7h
A+	85–90%	12,229	12.9h

How This Dataset Was Built

1. Source: Thai-language audio from sarulab-speech/yodas2_sidon (YouTube-sourced, CC-BY-3.0)
2. Transcription: Each utterance transcribed independently by 3 ASR models — biodatlab/distill-whisper-th-large-v3, typhoon-ai/typhoon-whisper-large-v3, and typhoon-ai/llama-3-typhoon-v1.5-8b-audio-preview. You can find the result of each model in the metadata.parquet file.
3. Verification: If 2 or more asr model match with the original text then user the original text. if only 1 match then run Google Gemini 3.0 Flash selected or merged the best transcription per utterance
4. Quality grading: Tiered by average ASR agreement score S+, S, A+ any score lower is dropped.
5. Text normalization: English words translated to Thai phonetics using a 4,502-word dictionary; numbers converted to Thai words
6. Audio processing: Resampled to 24kHz mono PCM-16, silence trimmed, peak normalized to ~−1dB
7. Long clip splitting: Clips >15s split at the nearest silence pause; clips >20s with no pause dropped
8. Quality filtering: DNSMOS ≥ 3.0 (overall and background), duration 1.5–20s, text ≥ 10 characters

Audio Quality Note

DNSMOS mean scores are 3.33 (overall) and 4.10 (background). The overall score appears moderate because DNSMOS was trained on English speech and systematically underscores non-English audio. The high background score (4.10) confirms the audio is clean. All samples passed DNSMOS ≥ 3.0 on both overall and background thresholds.

Intended Use

This dataset was built for training Thai text-to-speech models. The multi-speaker coverage (~4,200 speakers) and consistent audio quality make it suitable for both single-speaker fine-tuning and multi-speaker TTS training. It may also be useful for Thai ASR fine-tuning.

Acknowledgements

This dataset would not exist without the original sarulab-speech/yodas2_sidon dataset. Many thanks to the YODAS2 team for making their data openly available under CC-BY-3.0.

License

CC-BY-3.0, consistent with the source YODAS2 dataset.

Citation

@misc{nakata2025sidonfastrobustopensource,
      title={Sidon: Fast and Robust Open-Source Multilingual Speech Restoration for Large-scale Dataset Cleansing}, 
      author={Wataru Nakata and Yuki Saito and Yota Ueda and Hiroshi Saruwatari},
      year={2025},
      eprint={2509.17052},
      archivePrefix={arXiv},
      primaryClass={cs.SD},
      url={[https://arxiv.org/abs/2509.17052](https://arxiv.org/abs/2509.17052)}, 
}

@inproceedings{li2023yodas,
  title={Yodas: Youtube-Oriented Dataset for Audio and Speech},
  author={Li, Xinjian and Takamichi, Shinnosuke and Saeki, Takaaki and Chen, William and Shiota, Sayaka and Watanabe, Shinji},
  booktitle={2023 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU)},
  pages={1--8},
  year={2023},
  organization={IEEE}
}