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:
mix.wav) — two overlapping speakers, optionally reverberant and noisys1.wav, s2.wav) — the individual speaker signalss1.txt, s2.txt) — Fun-ASR-Nano transcriptions of each target,
usable as recognition-loss targets or for overlapping-ASR research| 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 |
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
<subset>/metadata.jsonlOne 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>.jsonPer-shard synthesis manifests: a list of (mix path, s1 path, s2 path, length) entries
recording exactly what each of the four generation shards produced.
Both subsets are generated by the released pipeline (src/synth/) with fixed random seeds:
rir/, a
speaker-embedding gate (WeSpeaker cnceleb-resnet34-LM) additionally rejects pairs
with cosine similarity > 0.5; for pure/, speaker labels must differ.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.rir/), or inherited from
source-utterance transcripts (pure/).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.
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}
}
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.
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:
mix.wav) — two overlapping speakers, optionally reverberant and noisys1.wav, s2.wav) — the individual speaker signalss1.txt, s2.txt) — Fun-ASR-Nano transcriptions of each target,
usable as recognition-loss targets or for overlapping-ASR research| 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 |
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
<subset>/metadata.jsonlOne 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>.jsonPer-shard synthesis manifests: a list of (mix path, s1 path, s2 path, length) entries
recording exactly what each of the four generation shards produced.
Both subsets are generated by the released pipeline (src/synth/) with fixed random seeds:
rir/, a
speaker-embedding gate (WeSpeaker cnceleb-resnet34-LM) additionally rejects pairs
with cosine similarity > 0.5; for pure/, speaker labels must differ.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.rir/), or inherited from
source-utterance transcripts (pure/).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.
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}
}
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.