CoolFace
Datasetpublic

FILM6912/th-en-zh-tts-200k-enhanced

TH-EN-ZH Multi-speaker TTS Dataset (200K, RE-USE Enhanced) Speech-enhanced variant of FILM6912/th-en-zh-tts-200k. Every clip has been processed through NVIDIA RE-USE (universal speech enhancement, SEMamba) at its native sample rate, then re-encoded losslessly as FLAC (PCM_16). Same schema, same row order, same 200,000 rows (th 100k / en 50k / zh 50k): Column Type Description text string Transcript (identical to the original dataset) audio Audio Enhanced audio, FLAC… See the full description on the dataset page: https://huggingface.co/datasets/FILM6912/th-en-zh-tts-200k-enhanced.

sourceHugging Facecc0-1.0updated 2d agoView on Hugging Face
0likes438downloads
Dataset Card

TH-EN-ZH Multi-speaker TTS Dataset (200K, RE-USE Enhanced)

Speech-enhanced variant of FILM6912/th-en-zh-tts-200k. Every clip has been processed through NVIDIA [RE-USE](https://huggingface.co/nvidia/RE-USE) (universal speech enhancement, SEMamba) at its native sample rate, then re-encoded losslessly as FLAC (PCM_16).

Same schema, same row order, same 200,000 rows (th 100k / en 50k / zh 50k):

ColumnTypeDescription
textstringTranscript (identical to the original dataset)
audioAudioEnhanced audio, FLAC bytes at the source sample rate
speaker_idstringSource-prefixed speaker ID (cv_th_*, libritts_*, aishell3_*)
langstringth / en / zh

Why an enhanced variant?

The source corpora contain real-world noise/reverb/codec artifacts. RE-USE suppresses additive noise and reverberation while preserving speaker identity and linguistic content, giving cleaner TTS training targets. Enhancement was applied at native sample rates (24/32/44.1/48 kHz) — no resampling, no bandwidth extension.

Processing

  • —Model: nvidia/RE-USE (30-layer TF-Mamba SEMamba, 9.61M params)
  • —bf16 autocast on a single RTX 3060; enhancement at native sample rate per clip
  • —Magnitude masking + zero-portion sweep-artifact removal exactly as NVIDIA's reference inference.py
  • —Output re-encoded as FLAC (lossless vs the enhanced float waveform)
  • —Mean waveform correlation (enhanced vs original): ~0.91 across languages; RMS reduced ~10–15% (noise suppression)

Selection / filtering

Identical to the original dataset — see its README for source attribution and licenses (CC0 / CC-BY-4.0 / Apache-2.0).

Usage

python
from datasets import load_dataset
ds = load_dataset("FILM6912/th-en-zh-tts-200k-enhanced", split="train", streaming=True)
row = next(iter(ds))
print(row["lang"], row["speaker_id"], row["text"], row["audio"]["sampling_rate"])

Provenance

  • —Base dataset built 2026-09-19 (seed 20260919) from Common Voice 17.0 (th), LibriTTS-R (en), AISHELL-3 (zh).
  • —Enhancement run: enhance_dataset.py, resumed-safe, per-shard stats in final_stats.json.
  • —RE-USE is licensed under the NVIDIA One-Way Noncommercial License (NSCLv1) — this dataset therefore inherits noncommercial-use constraints for the enhanced audio. The original recordings remain under their source licenses.

<!-- ml-intern-provenance -->

Generated by ML Intern

This dataset repository was generated by ML Intern, an agent for machine learning research and development on the Hugging Face Hub.

  • —Try ML Intern: https://smolagents-ml-intern.hf.space
  • —Source code: https://github.com/huggingface/ml-intern