Refiner-DB

Refiner-DB is the speech refinement dataset associated with the LoopTTS paper (arXiv:2608.28970) and project repository. This release contains the processed training data: 42,338 examples with Gemini-generated natural-language prosody suggestions, plus a separate 200-example human-annotated test set. The earlier 42,420-example preliminary collection is not part of this release.

The training and test JSONL files retain full utterance text. Audio references are relative to the repository root; the corresponding audio is supplied in two archives. No machine-specific source paths are needed.

Files and splits

.
├── README.md
├── LICENSE
├── data_manifest.json
├── data/
│   ├── train.jsonl
│   ├── test.jsonl
│   ├── audio_paths.txt
│   └── test_audio_paths.txt
├── audio.tar.gz
└── test_audio.tar.gz

Split	Hugging Face config	Rows	Audio archive	Unique audio files
Gemini annotated training	gemini_train / train	42,338	audio.tar.gz	86,616
Human annotated test	human_test / test	200	test_audio.tar.gz	200

The two splits have different schemas and are exposed as separate Hugging Face configs. Their keys and audio files do not overlap.

To inspect the JSONL records with datasets:

from datasets import load_dataset

train = load_dataset("Poookeman/Refiner-DB", "gemini_train", split="train")
test = load_dataset("Poookeman/Refiner-DB", "human_test", split="test")

These configs load the annotations. Extract the audio archives as described below to open the referenced WAV files.

Audio setup

Download the repository and extract both archives at its root, beside data/:

hf download Poookeman/Refiner-DB --repo-type dataset --local-dir .
tar -xzf audio.tar.gz
tar -xzf test_audio.tar.gz

Both archives extract into the same audio/ directory. For example, Path(row["raw_wav"]) for a training row, or Path(row["audio_path"]) for a test row, resolves from the repository root after extraction. All paths in the JSONL and archive headers are relative. The archive files need to be extracted before the paths can be opened as local WAV files.

To fetch only the smaller human test set:

hf download Poookeman/Refiner-DB data/test.jsonl test_audio.tar.gz --repo-type dataset --local-dir .
tar -xzf test_audio.tar.gz

data/audio_paths.txt and data/test_audio_paths.txt list the distinct relative WAV paths used by each split. data_manifest.json provides row counts, archive sizes, and SHA-256 checksums.

Training records

Each line in data/train.jsonl is one JSON object with these fields:

Field	Meaning
key	Example identifier
source_text, target_text	Full utterance text
suggestion	Gemini-generated natural-language prosody and refinement instruction
raw_wav	Initial synthesized speech; relative WAV path
target_wav	Reference target speech; relative WAV path
neutral_speaker_wav	Neutral speaker reference; relative WAV path
raw_audio_token, refined_audio_token	Audio codec token sequences

The audio token numbers are not word indices for pauses or stress. Prosody cues for the training split are expressed in suggestion.

Training sources: LibriTTS (17,187), ESD (12,135), EmoVoice (11,238), RAVDESS (1,107), MESS (337), and SAVEE (334).

Human test records

Each line in data/test.jsonl is one JSON object with index, key, full-sentence text, emotion, relative audio_path, human stress and pause index lists, and note. The index lists were preserved from the selected human annotation export without renumbering. The test set has 184 EmoVoice and 16 LibriTTS examples.

License and sources

Refiner-DB is released under Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0). Attribution is required and commercial use is not permitted. See LICENSE for the full terms. The dataset is intended for noncommercial research.

The source corpora include LibriTTS, ESD, RAVDESS, SAVEE, and MESS. Please also acknowledge the applicable source corpus when using its audio.

For research using Refiner-DB, please cite LoopTTS using the citation in the project repository.