env-tts-sd-corpus

Environment-aware text-to-speech training corpus. Each row pairs three short 16 kHz mono FLAC clips with a transcript:

• an environment sample (different speaker, same acoustic scene),
• a speaker reference (same speaker, optionally with augmented acoustics),
• the target speech,

so a TTS model can learn to synthesise a target utterance with both a specified voice and a specified environment.

Schema

column	type	description
environment_audio_source	binary (FLAC 16 kHz mono)	acoustic-scene reference, ≤15 s, drawn from a different speaker than speech but the same recording session
environment_audio_duration	float32	seconds
speaker_audio_source	binary (FLAC 16 kHz mono)	speaker-identity reference, ≤15 s, from the same speaker as speech
speaker_audio_duration	float32	seconds
text	string	transcript of speech; either the original gold transcript or a fresh Qwen3-ASR re-label
speech	binary (FLAC 16 kHz mono)	target utterance, 3–15 s
speech_duration	float32	seconds
language	string	zh / en / auto
dataset	string	one of m3sd / aishell4 / msdwild / chime6
conversation_id	string	unique within the source dataset
speaker_id	string	within-conversation diarisation label
env_id	string	acoustic-scene identifier (usually the conversation_id)
text_source	string	original or asr
spk_aug	string	none / noise / rir+noise (only present when augmentation was applied)
spk_aug_snr_db	float32	signal-to-noise ratio used when spk_aug != none

Source corpora

dataset	hours	sessions	language	how we use it
M3SD (Wu et al., 2025)	770	1 372	zh / en mixed	YouTube speaker-diarisation corpus, multi-scenario, transcripts via Qwen3-ASR
AISHELL-4 (Fu et al., 2021)	120	211	zh	Mandarin meetings with native TextGrid transcripts
MSDWILD (Liu et al., 2022)	80	3 143	zh / en mixed	in-the-wild speaker-diarisation videos, transcripts via Qwen3-ASR
CHiME-6 (Watanabe et al., 2020)	40	18	en	dinner-party recordings (Kinect U06 / U01 binaural), official JSON transcripts

Processing pipeline

The pipeline is three streaming stages running in parallel as separate processes, with a small filesystem-based handoff for backpressure:

1. download — one thread per source. Hugging Face mirrors and direct tar URLs are streamed with httpx.stream; the tar bytes are never written to disk in full. Each upstream "conversation" emits a JSON sentinel under state/ready/ as soon as its audio is locally addressable.

2. process — an asyncio loop drains sentinels with a bounded semaphore (default 64 concurrent conversations). For each conversation it

   • resamples the audio to 16 kHz mono,
   • walks the diarisation turns, chunks each turn into 3–15 s pieces,
   • picks a same-speaker reference (≥3 s, concatenating short turns when needed) and a different-speaker environment slice (≥3 s, extended into surrounding audio if necessary),
   • submits any speech pieces whose transcript is missing or whose turn was split mid-utterance to a Qwen3-ASR-1.7B Flask worker for fresh ASR,
   • FLAC-encodes the three clips and appends a row to the sharded parquet writer.

   The ASR worker coalesces concurrent requests into length-bucketed batches ([0–4 s], [4–9 s], [9–16 s], 16 s+) so that the HuggingFace padding=True step inside qwen-asr does not waste GPU on long zero-pad tails. Single-clip OOMs are dropped silently (the row is dropped, not the sibling 255 clips).

3. upload — watches final/upload_queue/group_*/ for sealed groups and uploads them to this repo via HfApi.upload_folder. Each group bundles ≈3 200 rows (4 parquet shards × 800 rows). Commits are rate-limited.

The reader, ASR worker, augmenter, and writer are all designed to recover cleanly from SIGKILL: all state is captured in a few small JSON files under state/ and an atomic-rename .tmp → final write protocol for each parquet shard.

Provenance: row breakdown by dataset

Counts are approximate (depend on streaming end + final partial groups).

dataset	records	rows emitted
M3SD	1 372	≈212 000
MSDWILD	3 113	≈ 28 400
AISHELL-4	145	≈ 35 250
CHiME-6	18	≈ 15 800

Drop reasons that account for "records ingested" > "records emitted with rows" in MSDWILD/M3SD: conversations with <2 speakers (no candidate for the env source), conversations whose total speech time per speaker can't yield a ≥3 s focal clip + a ≥3 s reference, or clips where Qwen3-ASR returned empty text after segmentation.

Licensing

The derived corpus is released under CC-BY-SA-4.0, which inherits the most-restrictive licence among the four sources. Note in particular:

• M3SD is for academic and non-commercial research only (Wu et al., 2025).
• MSDWILD uses the X-LANCE research-only agreement (Liu et al., 2022).
• AISHELL-4 (Apache 2.0) and CHiME-6 (CC-BY-SA-4.0) are open.

If you redistribute audio extracted from this dataset, you must comply with M3SD's and MSDWILD's non-commercial restriction.

Citation

If you use this corpus, please cite the four source papers:

@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 was performed with Qwen3-ASR-1.7B.

Loading

from datasets import load_dataset

ds = load_dataset("ChristianYang/env-tts-sd-corpus", split="train", streaming=True)
row = next(iter(ds))
print(row["text"])
print(row["speech"]["sampling_rate"], len(row["speech"]["array"]))

The audio columns are typed as the HF Audio feature (16 kHz, mono), so they decode automatically on access.

Files on disk

data/
  group_00000/
    manifest.json
    data_000000.parquet
    data_000001.parquet
    data_000002.parquet
    data_000003.parquet
  group_00001/
    ...

Each group is one atomic HF commit. Each parquet shard is ≈800 rows; group size is 4 × 800 = 3 200 rows ≈ 250 MB (snappy-compressed, audio columns already FLAC).

---

Source pipeline: https://github.com/... (see the linked repository for the streaming download/process/upload code that produced this dataset).