FineEdit: A Structured Paired Dataset for Controllable TTS

FineEdit is a large-scale, structured paired English speech dataset constructed to support FineCombo-TTS (Interspeech 2026). It is specifically designed for learning precise relative acoustic attribute control in text-to-speech synthesis.

Each sample is organized as a triplet: (source speech, control description, target speech). Since the source and target speech differ in only one specific acoustic attribute, it allows the model to learn precise reference-grounded transformations rather than absolute properties.

---

📁 Directory Structure & Sources

• pair-prosody/: JSON metadata for relative prosody control. Built on LibriTTS-R with speech rate and pitch modifications.
• pair-emotion/: JSON metadata for relative emotion control. Built on the ESD dataset across 5 emotions.
• pair-timbre/: JSON metadata for relative timbre control. Built by cross-pairing speakers with similar acoustic profiles using LibriTTS-P and TextrolSpeech annotations.
• code-change-prosody/: Parallelized batch processing scripts (FFmpeg + GNU Parallel) used for prosody manipulation.
• metadata/: Contains the baseline meta-information and speaking style/prosody annotations directly derived from LibriTTS-R and TextrolSpeech.

---

📥 Data Download

Due to GitHub's file size limits, the large augmented audio files are hosted on Google Drive. The text annotations and scripts are available directly in this repository. For your convenience, the official links to our source datasets are also provided below.

1. FineEdit Components

Component	Format	Source	Download Link
Paired Metadata	.json	GitHub	Repo
Augmentation Scripts	.sh	GitHub	code-change-prosody/
Prosody Audio Pack	.tar.gz	Google Drive	👉 Download from Google Drive

2. Upstream Source Datasets

To fully utilize the metadata or replicate the full data environment, you can access the original source datasets here:

• LibriTTS-R: https://www.openslr.org/141/
• LibriTTS-P: https://github.com/line/LibriTTS-P
• TextrolSpeech: https://github.com/jishengpeng/TextrolSpeech

---

📝 Citation

If you find this dataset or project helpful, please cite our Interspeech 2026 paper:

@inproceedings{zhou2026finecombo,
  title={FineCombo-TTS: Collaborative and Precise Controllable Speech Synthesis Using Text Descriptions and Reference Speech},
  author={Zhou, Shuoyi and Zhou, Yixuan and Yang, Peiji and, Hu, Yifan and Zhong, Yicheng and Wang, Zhisheng and Wu, Zhiyong},
  booktitle={Interspeech},
  year={2026}
}