README.md

---
license: cc-by-nc-4.0
language:
- en
- zh
tags:
- env-tts
- environment-aware-tts
- speaker-diarization
- text-to-speech
- audio
- speech-synthesis
size_categories:
- 100K<n<1M
pretty_name: Env-TTS-Clean
configs:
- config_name: default
  data_files:
  - split: train
    path: data/group_*/*.parquet
  - split: test
    path: test/data-*.parquet
  - split: validation
    path: validation/data-*.parquet
---

# Env-TTS-Clean

**Environment-aware text-to-speech training corpus (clean release).** Each row
pairs four short **24 kHz mono FLAC** clips with aligned transcripts:

- an **environment** sample (different speaker, same acoustic scene),
- a **speaker** reference (same speaker as the target utterance),
- a **speaker-enhanced** copy of the reference (MossFormer2 enhancement — or, for
  the **DDS** source, the **real clean-studio recording** of the speaker reference),
- the target **speech** to synthesise,

so a model can learn to generate an utterance with both a specified voice and a
specified environment. This release supersedes the earlier
[Env-TTS-SD-Corpus](https://huggingface.co/datasets/humanify/Env-TTS-SD-Corpus)
with a richer schema, higher sample rate, per-clip ASR for all three contexts,
two meeting sources (AMI, AliMeeting), and a controlled read-speech source
(**DDS / DAPS**) that contributes real clean-studio enhancement ground truth.

## Dataset statistics

| metric | value |
| --- | ---: |
| **rows** | **384,318** (189,918 diarization + 194,400 DDS) |
| **on-disk size** | ~197 GB |
| **Σ `speech_duration`** | **~680 h** (428.6 h diarization + ~251.7 h DDS) |
| **Σ (`environment_audio_duration` + `speaker_audio_duration` + `speech_duration`)** | **~1,964 h** (1,252.1 h + ~712 h DDS) |

Diarization-source totals were computed exactly on 2026-05-27 (239 shards). The
**DDS** portion (194,400 rows, added 2026-06-09; train groups `group_00061`–
`group_00122`) is estimated from a sampled shard (mean per row: speech 4.66 s,
env 4.24 s, speaker 4.29 s).

### Rows by source dataset

| `dataset` | rows |
| --- | ---: |
| `dds` | 194,400 |
| `m3sd` | 108,708 |
| `aishell4` | 27,924 |
| `alimeeting` | 25,087 |
| `ami` | 13,207 |
| `msdwild` | 9,676 |
| `chime6` | 5,316 |

## Schema

| column | type | description |
| --- | --- | --- |
| `environment_audio_source` | binary (FLAC 24 kHz mono) | acoustic-scene reference, 2.5–15 s, from a **different speaker** in the same scene |
| `environment_audio_duration` | float32 | seconds |
| `environment_audio_text` | string | transcript of the environment clip (gold / Qwen3-ASR / DAPS script) |
| `speaker_audio_source` | binary (FLAC 24 kHz mono) | speaker-identity reference, 2.5–15 s, **same speaker** as `speech` |
| `speaker_audio_duration` | float32 | seconds |
| `speaker_audio_text` | string | transcript of the speaker reference clip |
| `speaker_audio_source_enhanced` | binary (FLAC 24 kHz mono) | de-environment'd speaker reference: **MossFormer2-enhanced** for diarization sources; the **real clean-studio recording** for DDS |
| `text` | string | transcript of `speech` |
| `speech` | binary (FLAC 24 kHz mono) | target utterance, 3–15 s |
| `speech_duration` | float32 | seconds |
| `language` | string | `zh` / `en` / `auto` |
| `dataset` | string | `dds` / `m3sd` / `aishell4` / `msdwild` / `chime6` / `ami` / `alimeeting` |
| `conversation_id` | string | unique within the source dataset (for DDS: `{room}__{device}__{channel}`) |
| `speaker_id` | string | within-scene speaker label (for DDS: DAPS speaker, e.g. `f1`, `m8`) |
| `env_id` | string | acoustic-scene identifier (for DDS: `dds__{room}__{device}__{channel}`) |
| `text_source` | string | `original`, `asr`, or `mixed` |
| `asr_token_count` | int32 | Qwen3-ASR token count for `speech` (0 when `text_source=original`) |
| `asr_mean_logprob` | float32 | mean log-prob per token for `speech` |

## Source corpora

| dataset | hours (≈, this release) | language | transcripts |
| --- | ---: | --- | --- |
| **DDS — Device-Degraded Speech, DAPS portion** (Li & Yamagishi, 2021) | ~252 | en | ✅ DAPS scripts |
| **M3SD** (Wu et al., 2025) | 770 | zh / en mixed | ❌ → Qwen3-ASR |
| **AISHELL-4** (Fu et al., 2021) | 120 | zh | ✅ TextGrid |
| **MSDWILD** (Liu et al., 2022) | 80 | zh / en mixed | ❌ → Qwen3-ASR |
| **CHiME-6** (Watanabe et al., 2020) | 40+ | en | ✅ JSON |
| **AMI** (SDM, diarizers-community) | ~100 | en | ❌ → Qwen3-ASR |
| **AliMeeting** (OpenSLR 119, far ch.0) | ~120 | zh | ✅ TextGrid |

**DDS** is single-speaker **read speech** (not a diarization corpus): 20 DAPS
speakers re-recorded across **9 rooms × 3 microphones × 6 positions** (162
acoustic conditions). For each condition the target `speech` and the
`environment_audio_source` (a *different* speaker, same room/mic/position) are the
**device-degraded** recordings, while `speaker_audio_source_enhanced` is the
matching **clean-studio** recording — a real enhancement ground truth rather than
a MossFormer2 estimate. All DDS text comes from the DAPS scripts
(`text_source = original`).

## Processing pipeline

Built with the streaming pipeline in
[`env-tts-data-pipeline`](https://github.com/ChristianYang/env-tts-data-pipeline).

Diarization sources — three parallel stages **download → process → upload**:

1. **download** — stream each source conversation (HF mirrors, OpenSLR tar
   streams, etc.) into a bounded local cache; emit a JSON sentinel when ready.
2. **process** — resample to **24 kHz mono**, walk diarisation turns, emit
   3–15 s `speech` slices with a same-speaker reference (≥2.5 s) and a
   different-speaker environment slice (≥2.5 s). Missing/split transcripts are
   re-labelled with **Qwen3-ASR-1.7B**. Snappy parquet shards (~800 rows / shard,
   4 shards per HF commit group).
3. **upload** — `HfApi.upload_folder` per sealed group, resume-safe.
4. **enhance** (second pass) — **MossFormer2_SE_48K** on `speaker_audio_source`.

**DDS** uses a dedicated parallel-channel pass (`process-dds`): each (room,
device, position) condition is one acoustic scene; rows are assembled directly
from the parallel clean/degraded recordings, and `speaker_audio_source_enhanced`
is filled in-place with the real clean-studio clip (no second-pass MossFormer2).

## Licensing

Released under **CC-BY-NC-4.0** (non-commercial), inheriting the most restrictive
terms among sources. In particular:

- **DDS / DAPS** — CC-BY-NC-4.0 (non-commercial). *This is the binding term for
  the whole release.*
- **M3SD** — academic / non-commercial research only.
- **MSDWILD** — X-LANCE research-only agreement.
- AISHELL-4 (Apache-2.0), CHiME-6 (CC-BY-SA-4.0), AMI, and AliMeeting carry
  their respective open / research terms.

Redistributing extracted audio requires complying with each upstream licence.

## Citation

Please cite the source papers when using this corpus:

```bibtex
@article{li2021dds,
  title={DDS: A new device-degraded speech dataset for speech enhancement},
  author={Li, Haoyu and Yamagishi, Junichi},
  journal={arXiv preprint arXiv:2109.07931},
  year={2021}
}

@article{wu2025m3sd,
  title={M3SD: Multi-modal, Multi-scenario and Multi-language Speaker
         Diarization Dataset},
  author={Wu, Shilong and others},
  journal={arXiv preprint arXiv:2506.14427},
  year={2025}
}

@inproceedings{fu2021aishell4,
  title={AISHELL-4: An Open Source Dataset for Speech Enhancement, Separation,
         Recognition and Speaker Diarization in Conference Scenario},
  author={Fu, Yihui and others},
  booktitle={Interspeech},
  year={2021}
}

@inproceedings{liu2022msdwild,
  title={MSDWILD: Multi-modal Speaker Diarization Dataset in the Wild},
  author={Liu, Tao and others},
  booktitle={Interspeech},
  year={2022}
}

@inproceedings{watanabe2020chime6,
  title={CHiME-6 Challenge: Tackling Multispeaker Speech Recognition for
         Unsegmented Recordings},
  author={Watanabe, Shinji and others},
  booktitle={CHiME Workshop},
  year={2020}
}
```

ASR re-labelling uses [Qwen3-ASR-1.7B](https://huggingface.co/Qwen/Qwen3-ASR-1.7B).
Speaker enhancement uses MossFormer2 (ClearVoice). DDS is built on the
[DAPS](https://zenodo.org/records/4660670) dataset (Mysore, 2015).

## Loading

```python
from datasets import load_dataset

ds = load_dataset("ChristianYang/Env-TTS-Clean", split="train", streaming=True)
row = next(iter(ds))
print(row["text"], row["dataset"], row["speech_duration"])
# Audio columns decode automatically when accessed (24 kHz mono).

# Filter to a single source (e.g. the DDS read-speech rows):
dds = ds.filter(lambda r: r["dataset"] == "dds")
```

## Files on disk

```
data/
  group_00000/ ... group_00122/      # group_00061–00122 are DDS
    manifest.json
    data_000000.parquet
    ...
```

Each `group_*` directory is one atomic HF commit bundle (typically 4 × 800-row
parquet shards, snappy-compressed FLAC payloads inside).