REAL-PS4

Overview

REAL-PS4 is a training dataset for target speaker extraction (TSE) and overlapping speech recognition in real-world multi-speaker meetings. It is constructed from four public meeting corpora — AISHELL-4, AliMeeting, AMI, and CHiME-6 — following the REAL-T benchmark data preparation pipeline.

Each sample consists of:

• A mixture utterance (overlapping speech clip from a real meeting)
• An enrollment utterance (clean speech from the target speaker in the same session)
• Ground-truth transcript of the target speaker
• Target activity segments (frame-level VAD labels for the target speaker)

---

Dataset Statistics

Subset	Language	Mixtures	Enrollment Clips	Meta Rows	Total Mix Duration
AISHELL-4	zh	1,083	4,295	11,465	~549 MB audio
AliMeeting	zh	3,332	1,760	33,148	~1.9 GB audio
AMI	en	1,550	2,020	14,933	~889 MB audio
CHiME-6	en	1,091	517	12,225	~496 MB audio
Total	zh / en	7,056	8,592	71,771	~5.9 GB

---

Directory Structure

REAL-PS4/
├── AISHELL-4/
│   ├── mapping.csv                         # mixture_utterance → wav path
│   ├── mixtures/                           # overlapping speech WAVs (16 kHz mono)
│   │   └── AISHELL-4_<session>_mixture_<start>_<end>.wav
│   ├── enrolment_speakers/                 # enrollment WAVs (16 kHz mono)
│   │   └── AISHELL-4_<session>_<speaker>_<start>_<end>.wav
│   └── TRAIN/
│       ├── AISHELL-4_meta.csv              # main metadata
│       ├── target_activity_segments.jsonl  # per-speaker VAD labels
│       └── json/AISHELL-4/
│           ├── overlap_records.json
│           ├── no_overlap_segments.json
│           └── overlap_statistics.json
├── AliMeeting/  (same structure)
├── AMI/         (same structure)
└── CHiME6/      (same structure)

---

File Formats

TRAIN/<DATASET>_meta.csv

Main metadata file. Each row represents one (mixture, enrollment, target speaker) triple.

Column	Description
mixture_utterance	Unique ID of the mixture clip
enrolment_speakers_utterance	Unique ID of the enrollment clip
source	Source dataset name
language	zh or en
total_number_of_speaker	Number of speakers in the mixture
speaker	Target speaker ID
gender	Speaker gender (M / F / U)
speaker_ratio	Fraction of mixture duration the target speaker is active
mixture_ratio	Fraction of mixture that contains overlapping speech
enrolment_speakers_duration	Duration of enrollment clip (seconds)
mixture_overlap_duration	Total overlapping duration in mixture (seconds)
mixture_duration	Total mixture duration (seconds)
ground_truth_transcript	Reference transcript of the target speaker

TRAIN/target_activity_segments.jsonl

JSONL file. Each line is a JSON object with target speaker VAD labels in local time (relative to mixture start):

{
  "mixture_utterance": "AISHELL-4_..._mixture_987.48_1004.59",
  "speaker": "005-F",
  "mixture_start": 987.48,
  "mixture_end": 1004.59,
  "mixture_duration": 17.11,
  "segments": [[0.0, 8.53], [11.27, 14.75]],
  "speech_ratio": 0.7025,
  "num_segments": 2,
  "source": "AISHELL-4"
}

TRAIN/json/<DATASET>/overlap_records.json

Per-session overlap region records, including all speakers and their ratios within each overlapping window.

mapping.csv

Maps mixture_utterance IDs to their WAV file paths on disk.

---

Audio Format

• Sample rate: 16,000 Hz
• Channels: Mono
• Format: WAV (PCM 16-bit)
• Amplitude-normalized to peak 0.95

---

Data Construction

Data is constructed from the raw corpora using the following pipeline:

1. Parse speaker diarization annotations (RTTM / TextGrid / segments XML / transcription JSON)
2. Extract non-overlapping segments as enrollment candidates
3. Detect and merge overlapping regions (minimum overlap: 5 s, mix duration: 5–100 s)
4. Filter by target speaker ratio (≥ 20%) and transcript length (≥ 5 characters/words)
5. Generate VAD labels (local time coordinates) from raw annotations

---

Source Datasets

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

Source	License	Description
AISHELL-4	CC BY-SA 4.0	Chinese meeting corpus (8-mic)
AliMeeting	CC BY-SA 4.0	Chinese multi-person meetings (far-field)
AMI Corpus	CC BY 4.0	English meeting recordings
CHiME-6	Research use	English dinner party recordings

---

Citation

If you use REAL-PS4 in your research, please cite:

@article{ning2026ps4,
  title     = {PS4: Proxy-Supervised Joint Training for Real Target Speaker Extraction},
  author    = {Wanyi Ning, Wei Zhou, Yingpeng Li, Yinshang Guo, Haitao Qian, Yiming Cheng},
  year      = {2026},
  publisher = {Arxiv}
}

---

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.