DAVE-Corpus (Open Subset)

Overview

DAVE-Corpus (open subset) is a training dataset for blind source separation (BSS) and two-speaker speech separation on Chinese meeting speech. It is the redistributable portion of the training pool of DAVE (arXiv:2608.09288), our system for the ISCSLP 2026 Real-World AVSE Challenge, and is generated end-to-end by the released synthesis pipeline from three permissively licensed corpora — AliMeeting, AISHELL-4 (speech) and MUSAN (noise).

Each pair consists of:

• A mixture utterance (mix.wav) — two overlapping speakers, optionally reverberant and noisy
• Two target references (s1.wav, s2.wav) — the individual speaker signals
• Two transcripts (s1.txt, s2.txt) — Fun-ASR-Nano transcriptions of each target, usable as recognition-loss targets or for overlapping-ASR research

---

Dataset Statistics

Subset	Pairs	Audio Files	Mixture Duration	Audio Size	Recipe
rir/	8,000	24,000	~13.3 h	~4.6 GB	50% direct sum + 50% simulated-room RIR (RT60 0.2-0.4 s), MUSAN noise at 12-30 dB SNR
pure/	4,000	12,000	~6.6 h	~2.3 GB	direct sum of RMS-normalized utterances, exact-zero target residual
Total	12,000	36,000	~19.9 h	~6.9 GB	4-8 s (rir/) / 2-6 s (pure/) per pair

---

Directory Structure

DAVE-Corpus/
├── rir/
│   ├── metadata.jsonl          # dataset-viewer index (file_name, pair, role, transcript)
│   ├── manifest_{0..3}.json    # synthesis-shard manifests (output paths + lengths)
│   └── rir_<shard>_<index>/    # one pair per directory
│       ├── mix.wav             # two-speaker mixture (+ reverb / noise)
│       ├── s1.wav  s2.wav      # target references
│       └── s1.txt  s2.txt      # Fun-ASR-Nano transcripts
└── pure/
    ├── metadata.jsonl
    └── pure_<index>/           # same five-file layout, direct-sum recipe

---

File Formats

<subset>/metadata.jsonl

One row per audio file; drives the dataset viewer and makes transcripts loadable as columns.

Column	Description
file_name	Path of the audio file relative to the subset root
pair	Pair directory name (groups mix / s1 / s2 of one mixture)
role	mixture, source1, or source2
transcript	Fun-ASR-Nano transcription (empty for mixture rows)

rir/manifest_<shard>.json

Per-shard synthesis manifests: a list of (mix path, s1 path, s2 path, length) entries recording exactly what each of the four generation shards produced.

---

Audio Format

• Sample rate: 16,000 Hz
• Channels: Mono
• Format: WAV (PCM 16-bit)
• Mixture and targets jointly peak-normalized to <= 0.99

---

Data Construction

Both subsets are generated by the released pipeline (src/synth/) with fixed random seeds:

1. Source pool — 1,505 single-speaker utterance segments drawn from the AliMeeting and AISHELL-4 training sets.
2. Pair sampling — two utterances from different speakers. For rir/, a speaker-embedding gate (WeSpeaker cnceleb-resnet34-LM) additionally rejects pairs with cosine similarity > 0.5; for pure/, speaker labels must differ.
3. Mixing — rir/ (4-8 s): RMS-normalize; with probability 0.5 convolve both sources with pyroomacoustics room impulse responses (shoebox rooms 3-8 m x 3-7 m x 2.5-3.5 m, RT60 0.2-0.4 s, random microphone / source positions), otherwise sum directly; scale the second speaker by a +-3 dB relative gain; add MUSAN noise at an SNR drawn from 12-30 dB. pure/ (2-6 s): direct sum with +-3 dB relative gain, no reverberation, no noise.
4. Post-processing — joint peak normalization to <= 0.99, written as 16 kHz mono PCM16.
5. Transcripts — Fun-ASR-Nano over each clean stream (rir/), or inherited from source-utterance transcripts (pure/).

---

Source Datasets

This dataset is derived from the following corpora (not redistributed here — only synthesized derivatives):

Source	License	Description
AliMeeting	CC BY-SA 4.0	Chinese multi-person meetings (far-field)
AISHELL-4	CC BY-SA 4.0	Chinese meeting corpus (8-mic)
MUSAN	CC BY 4.0	Music, speech and noise corpus (noise portion used)

Deliberately excluded: no audio in this dataset derives from MISP-Meeting (non-commercial agreement) or from ISCSLP 2026 challenge development data (challenge terms). The paper's full training pool additionally contained subsets built from those sources; they cannot be redistributed. To rebuild them, download the corpora under their own licenses and run the released pipeline — see DATA_LICENSE.

---

Citation

If you use DAVE-Corpus in your research, please cite:

@misc{zhou2026dave,
  title         = {{DAVE}: A Decoupled Audio-Visual Enhancement Framework for Real-World Speech Separation},
  author        = {Zhou, Wei and Ning, Wanyi and Guo, Yinshang and Fang, Qianxiao and Qian, Haitao and Li, Yingpeng},
  year          = {2026},
  eprint        = {2608.09288},
  archivePrefix = {arXiv},
  primaryClass  = {cs.SD},
  url           = {https://arxiv.org/abs/2608.09288}
}

---

License

This dataset is released under the Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0) license, inherited from its CC BY-SA source corpora, with attribution to their creators.