README.md

---
pretty_name: Bagpiper-TTS SFT Data
language:
- en
task_categories:
- text-to-speech
- text-to-audio
tags:
- audio
- speech
- singing
- text-to-speech
- multimodal
- parquet
size_categories:
- 100K<n<1M
---

# Bagpiper-TTS SFT Data

> **Release status:** the validated Parquet release is being uploaded. The
> homepage and metadata may appear before every large shard is committed.

Bagpiper-TTS SFT Data supports
[Bagpiper-TTS](https://bagpipertts.github.io/bagpiper_tts_demo/), a universal
speech-synthesis model that interprets free-form natural-language requests,
plans the requested delivery, produces a rich textual caption, and synthesizes
the target audio.

The release is organized into the six applications used by the paper:

| Configuration | Capability |
|---|---|
| `classical_tts` | Explicit text with natural-language voice and delivery instructions |
| `multi_talker` | Multi-speaker dialogue with per-speaker characteristics |
| `intent_to_speech` | Communicative intent without requiring exact wording |
| `role_play` | Persona/character-driven expressive speech |
| `svs` | Singing voice synthesis with lyrics and musical instructions |
| `general_purpose` | Open-ended speech scenes and non-standard requests |

Individual source corpora remain provenance fields and do not create extra
public configurations.

## Loading the data

Named Hugging Face configurations are deferred. Load an application with its
explicit Parquet glob; streaming is recommended:

```python
from datasets import load_dataset

dataset = load_dataset(
    "parquet",
    data_files={"train": "hf://datasets/espnet/Bagpiper_TTS_SFT_Data/intent_to_speech/*.parquet"},
    split="train",
    streaming=True,
)

example = next(iter(dataset))
audio_bytes = example["audio"]["bytes"]
```

Every example will be self-contained: its request, planning/caption text, and
complete encoded audio bytes are stored in the same Parquet row. No external
audio lookup or table join will be required. Data will be split across many
deterministic Parquet shards targeting approximately 256 MiB each.

The published training conversation intentionally contains no system message:
each row is ordered as `user:text → assistant:text → assistant:audio`, matching
the paper's no-system-prompt description. The leading task-specific system
messages found in raw staging JSONLs were removed from all 738,123 selected
rows by owner decision.

See [Schema](docs/SCHEMA.md) and
[Release status](docs/RELEASE_STATUS.md).

## Data construction

The pipeline starts from curated speech or singing audio, creates a detailed
rich caption, extracts or verifies the transcription, filters transcription
errors, reverse-simulates varied natural-language requests, constructs a
three-part planning trace, and applies consistency filtering. The paper uses
Qwen3-235B-A22B-Instruct-FP8 as its primary text processor and optionally uses
Gemini audio validation.

## Release statistics

| Application partition | Rows | Shards | Parquet bytes | Embedded audio bytes |
|---|---:|---:|---:|---:|
| `classical_tts` | 235,279 | 241 | 52,899,149,925 | 57,191,194,377 |
| `multi_talker` | 64,659 | 170 | 38,632,502,383 | 44,989,933,724 |
| `intent_to_speech` | 153,568 | 325 | 68,640,618,698 | 85,622,064,832 |
| `role_play` | 47,116 | 107 | 22,456,138,107 | 28,130,403,188 |
| `svs` | 101,887 | 120 | 29,503,615,421 | 31,274,125,012 |
| `general_purpose` | 135,614 | 188 | 40,849,052,445 | 48,322,109,074 |
| **Total** | **738,123** | **1,151** | **252,981,076,979** | **295,529,830,207** |

`classical_tts` contains 17,656 LibriTTS-R clean-100, 60,657 clean-360,
103,267 other-500, 40,793 Genshin, and 12,906 Star Rail rows. Intended reuse
of valid audio across rows is retained.

## Intended use and non-use

The data is intended for research on instruction-following speech synthesis.
Bagpiper-TTS uses textual requests and does not accept reference audio; this
release should not be described as a voice-cloning dataset or system.

## Known limitations

- Rich captions and simulated planning can hallucinate attributes or content.
- The six applications have heterogeneous sources and filtering rules.
- Some raw source collections have unresolved redistribution terms. Affected
  examples will be withheld unless rights are confirmed.
- Game-derived voices, singing material, and benchmark/evaluation assets
  require particular care; public availability of a raw staging repository is
  not itself redistribution permission.
- The dataset may contain synthetic or transformed audio and machine-generated
  text rather than expert annotation.
- The corpus is not exhaustively moderated for personal information, unsafe
  content, offensive language, or copyrighted text/lyrics.

## Related resources

- Project/demo: [Bagpiper-TTS](https://bagpipertts.github.io/bagpiper_tts_demo/)
- Bagpiper paper: [OpenReview](https://openreview.net/forum?id=FuHs64E3X6)
- Bagpiper project: [bagpiper-cmu.github.io](https://bagpiper-cmu.github.io/)
- Base model: [espnet/bagpiper](https://huggingface.co/espnet/bagpiper)
- ESPnet: [espnet/espnet](https://github.com/espnet/espnet)

## Citation

Please cite the Bagpiper-TTS paper and Bagpiper foundation-model paper when
using this dataset. Copy-ready citation metadata will be added once the final
Bagpiper-TTS venue/arXiv record is confirmed.

The accepted Bagpiper foundation-model citation is:

```bibtex
@inproceedings{anonymous2026bagpiper,
  title={Bagpiper: Solving Open-Ended Audio Tasks via Rich Captions},
  author={Jinchuan Tian and Haoran Wang and Bo-Hao Su and Chien-yu Huang and
          Qingzheng Wang and Jiatong Shi and William Chen and Xun Gong and
          Siddhant Arora and Chin-Jou Li and Masao Someki and Takashi Maekaku and
          Keita Goto and Yusuke Shinohara and Jin Sakuma and
          Chao-Han Huck Yang and Shinji Watanabe},
  booktitle={Third Conference on Language Modeling},
  year={2026},
  url={https://openreview.net/forum?id=FuHs64E3X6}
}
```

## Contact and takedown

Please use the repository community tab for provenance corrections or takedown
requests. Include the configuration and `example_id`; do not repost sensitive
media in the report.