README.md

---
license: mit
---

# b150_official_train

This release contains the high-confidence `b150` training split used by MAGIC-TTS,
with standalone MFA word-level alignments distributed separately from prepared Arrow
artifacts.

Related links:
- GitHub repo: https://github.com/yongaifadian1/MAGIC-TTS
- arXiv paper: https://arxiv.org/abs/2604.21164
- Open-source model: https://huggingface.co/maimai11/MAGIC-TTS
- Online demo: https://yongaifadian1.github.io/MAGIC-TTS/

Stats:
- Total samples: `201986`
- English: `78363`
- Chinese: `123623`

Files:
- `selected_samples.exported.jsonl`: manifest with `sample_id`, `utt_id`, `language`, `audio_path`, `duration`, and `target_text`
- `raw_audio_en.tar.part-*`: split English raw audio bundle
- `raw_audio_zh.tar.part-*`: split Chinese raw audio bundle
- `mfa_sidecars.tar.part-*`: split standalone MFA word-level alignments with relative `audio_path`, ready to be reused during local dataset preparation

Notes:
- Reconstruct tar files first:

```bash
cat raw_audio_en.tar.part-* > raw_audio_en.tar
cat raw_audio_zh.tar.part-* > raw_audio_zh.tar
cat mfa_sidecars.tar.part-* > mfa_sidecars.tar
```

- Then extract:

```bash
mkdir -p raw
tar -xf raw_audio_en.tar -C raw
tar -xf raw_audio_zh.tar -C raw
tar -xf mfa_sidecars.tar
```

- After extraction, the dataset layout becomes `raw/audio/...` and `mfa_sidecars/...`.
- Arrow artifacts are intentionally not included in this release.