README.md

---
pretty_name: REAL-T
license: cc-by-sa-4.0
language:
  - en
  - zh
task_categories:
  - audio-to-audio
tags:
  - audio
  - speech
  - target-speaker-extraction
size_categories:
  - 1K<n<10K
---

# REAL-T: Real Conversational Mixtures for Target Speaker Extraction

REAL-T is a conversation-centric benchmark for **target speaker extraction (TSE)** in real meetings and dinner-party recordings. Unlike simulated stacks such as LibriMix, the mixtures are cut from naturally overlapping speech, with enrollment clips taken from non-overlapping regions of the same conversations.

This package contains the **DEV**, **EVAL1**, and **EVAL2** splits used by the [REAL-TSE Challenge](https://real-tse.github.io/challenge/) (a satellite challenge of IEEE SLT 2026). Official inference and scoring live in [REAL-TSE-Challenge](https://github.com/REAL-TSE/REAL-TSE-Challenge). The dataset paper is [Li et al., Interspeech 2025](https://www.isca-archive.org/interspeech_2025/li25da_interspeech.pdf) and [Wang et al., arXiv 2026](https://arxiv.org/pdf/2607.15198).

**Task.** Given a multi-speaker mixture and one or more enrollment utterances of a designated target speaker, recover that speaker’s speech. Isolated clean references are **not** provided (they do not exist for these real recordings). The official metrics are TER, speaker similarity, DNSMOS, and timing F1.

To run the released toolkit, copy the three split folders into the [challenge repo](https://github.com/REAL-TSE/REAL-TSE-Challenge):

```bash
cp -r REAL-T-dev REAL-T-eval1 REAL-T-eval2 /path/to/REAL-TSE-Challenge/datasets/
```

Then follow that repo (`pre.sh`, `run_tse.sh`, `run_eval.sh`). Local scoring with ground-truth transcripts is supported for **DEV** only.

License: [CC BY-SA 4.0](https://creativecommons.org/licenses/by-sa/4.0/). See [License and terms of use](#license-and-terms-of-use). 

**Release:** v1.0.0 (2026-08-23).

## Data layout

Each `*_meta.csv` row is one **sample**: one mixture + one enrollment of the target speaker, i.e. one evaluation trial. Several samples may reuse the same mixture or enrollment WAV, so the file counts in the tree below are smaller than the sample counts.

Duration and ratio cells are **sample-level means**.


| Statistic               | Meaning                                                                        | DEV                                                                  | EVAL1                                                | EVAL2                                   |
| ----------------------- | ------------------------------------------------------------------------------ | -------------------------------------------------------------------- | ---------------------------------------------------- | --------------------------------------- |
| Samples                 | Mix–enroll pairs (`*_meta.csv` rows); each pair is one TSE trial               | 1,991                                                                | 2,000                                                | 3,000                                   |
| Target speakers         | Distinct target-speaker IDs among those samples                                | 64                                                                   | 49                                                   | 77                                      |
| Language (zh / en)      | Sample counts in Mandarin vs English                                           | 721 / 1,270                                                          | 298 / 1,702                                          | 1,487 / 1,513                           |
| Mix duration (s)        | Mean length of the mixture clip                                                | 17.26                                                                | 18.23                                                | 17.25                                   |
| Enrollment duration (s) | Mean length of the target-speaker enrollment cue                               | 9.65                                                                 | 10.95                                                | 9.10                                    |
| Overlap ratio           | Mean fraction of the mixture where **two or more speakers talk at once**       | 0.49                                                                 | 0.48                                                 | 0.53                                    |
| Target ratio            | Mean fraction of the mixture where the **designated target speaker is active** | 0.75                                                                 | 0.74                                                 | 0.73                                    |
| Sources                 | Source corpora (sample count)                                                  | AISHELL-4 (240) AliMeeting (481) AMI (592) CHiME-6 (545) DiPCo (133) | AliMeeting (298) AMI (948) CHiME-6 (528) DiPCo (226) | `unseen_CN` (1,487) `unseen_EN` (1,513) |


These map to CSV fields `mixture_duration`, `enrolment_speakers_duration`, `mixture_ratio`, and `speaker_ratio`.

DEV uses the same **1,991 mix–enroll pairs** as the PRIMARY split in the Interspeech 2025 paper (same sample count and duration/ratio statistics). That is an inventory match, not a claim that the WAV files are bit-identical to a 2025 dump: the 2025 paper mixed specific microphone channels (first channel for AISHELL-4, AliMeeting, and AMI; the average of array channels for CHiME-6 and DiPCo), but the audio files of CHiME-6 & DiPCo samples in this package (SLT 2026 challenge release) are changed (because we re-selected channels and re-make the audio files).

EVAL1 is drawn from the same public corpora as DEV, with **no session overlap** and without AISHELL-4. AMI meeting IB4002 has a known, still-unfixed wrong headset channel mapping ([Data Problems](https://groups.inf.ed.ac.uk/ami/corpus/dataproblems.shtml): A→3, B→2, C→0, D→1) that left its RTTM speaker labels globally permuted — for example, audio labeled MIO091 actually contains FIE038. We noticed this unfixed error after the challenge, so the released dataset differs from the dataset at REAL-TSE Challenge (SLT 2026) time.

EVAL2 is newly collected Mandarin and English conversational audio (held-out rooms and devices), including matched- and cross-channel mix–enroll pairings. Scene, channel, and pairing-condition labels for each trial are in `EVAL2/scene_channel.csv` (see below). 

Relative to the REAL-TSE Challenge (SLT 2026) release, this package also re-normalizes every EVAL1 and EVAL2 `ground_truth_transcript` and the per-speaker / per-segment transcripts in the EVAL overlap JSON with the Whisper large-v2 tokenizer (English `normalize`, Chinese `basic_normalize`). DEV transcripts are left unchanged. As a result, **the annotations differ from the data used at challenge time**. Challenge-period EVAL1 and EVAL2 results are therefore not directly comparable to results computed on this release.

```text
REAL-T/
├── REAL-T-dev/
│   ├── mixtures/                      # 528 wav
│   ├── enrolment_speakers/            # 242 wav
│   └── DEV/
│       ├── AISHELL-4_meta.csv
│       ├── AliMeeting_meta.csv
│       ├── AMI_meta.csv
│       ├── CHiME6_meta.csv
│       ├── DipCo_meta.csv
│       └── json/<source>/overlap_records.json
├── REAL-T-eval1/                      # same layout; no AISHELL-4
│   ├── mixtures/                      # 595 wav
│   ├── enrolment_speakers/            # 183 wav
│   └── EVAL1/
└── REAL-T-eval2/
    ├── mixtures/                      # 1,186 wav
    ├── enrolment_speakers/            # 1,913 wav
    └── EVAL2/
        ├── unseen_CN_meta.csv
        ├── unseen_EN_meta.csv
        ├── scene_channel.csv          # scene, channel, and pairing conditions
        └── json/unseen_{CN,EN}/overlap_records.json
```

All audio is **16 kHz, mono, 16-bit PCM WAV**. The CSV fields `mixture_utterance` and `enrolment_speakers_utterance` are file stems (no `.wav`) under `mixtures/` and `enrolment_speakers/`. EVAL2 mixture stems encode the recording device as `H1`, `H2`, or `phone`. EVAL2 enrollment stems use those same three tokens, plus `A` for a speaker-worn headset and `external` for enrollments recorded outside the meeting. `mapping.csv` is not shipped; the challenge toolkit regenerates it with `pre.sh`.

Each `*_meta.csv` has one row per mix–enroll pair:


| Column                                                                        | Meaning                                                                                                |
| ----------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------ |
| `mixture_utterance`                                                           | Mixture file stem                                                                                      |
| `enrolment_speakers_utterance`                                                | Enrollment file stem                                                                                   |
| `source`                                                                      | Corpus or EVAL2 subset (`AISHELL-4`, `AliMeeting`, `AMI`, `CHiME6`, `DipCo`, `unseen_CN`, `unseen_EN`) |
| `language`                                                                    | `zh` or `en`                                                                                           |
| `total_number_of_speaker`                                                     | Speakers in the mixture                                                                                |
| `speaker`, `gender`                                                           | Target speaker id and gender (`M` / `F`)                                                               |
| `speaker_ratio`                                                               | Target speaker’s speaking proportion in the mixture                                                    |
| `mixture_ratio`                                                               | Overlap ratio of the mixture                                                                           |
| `enrolment_speakers_duration`, `mixture_overlap_duration`, `mixture_duration` | Durations in seconds                                                                                   |
| `ground_truth_transcript`                                                     | Normalized transcript of the **target** speaker in the mixture (used for TER)                          |




`json/<source>/overlap_records.json` stores per-speaker activity in each mixture. It is used for the timing F1 metric, not as training labels.

### EVAL2 scene and channel labels

`EVAL2/scene_channel.csv` adds recording-condition labels to the 3,000 EVAL2 trials (one row per mix–enroll pair). It joins to `unseen_CN_meta.csv` / `unseen_EN_meta.csv` on (`mixture_utterance`, `enrolment_speakers_utterance`) and does not repeat fields already in those files. Device names follow Table III of the [challenge paper](https://arxiv.org/abs/2607.15198).


| Column                        | Meaning                                                                                                                                                               |
| ----------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| `mixture_scene`               | Scene where the mixture was recorded: `open_office`, `large_meeting_room`, `medium_meeting_room`, `small_meeting_room`, `restaurant`, `home`, `in-vehicle`, or `cafe` |
| `mix_channel`                 | Mixture recording channel: `H1` (high-quality microphone), `H2` (high-quality microphone, farther away), or `phone`                                                   |
| `enrolment_kind`              | `direct` (enrollment cut from the same meeting session) or `external` (separate single-speaker recording made outside the meeting)                                    |
| `enrol_channel`               | Enrollment recording channel: `H1`, `H2`, `phone`, `headset` (speaker-worn close-talk), or `external`                                                                 |
| `enrol_scene`                 | Scene where the enrollment was recorded (same 8-value set as `mixture_scene`); empty for external enrollments                                                         |
| `scene_match`, `device_match` | `same` / `cross` between mixture and enrollment for direct enrollments; empty for external enrollments (not compared)                                                 |
| `quadrant`                    | `same_scene_same_device`, `same_scene_cross_device`, `cross_scene_same_device`, `cross_scene_cross_device`, or `external`                                             |


The 3,000 EVAL2 trials comprise 2,400 direct enrollments (600 per quadrant) and 600 external enrollments.

> **Note: these labels are provided for reference only.** As discussed in the challenge paper, per-scenario metric trends do not consistently follow intuitive acoustic expectations — meeting-room recordings are not necessarily easier, and restaurant-like scenarios are not necessarily harder. This is because EVAL2 conversations were recorded in a natural, topic-driven manner with free speaker interaction, so each scenario contains substantial variability in turn-taking, overlap, speaker distance, loudness, device placement, and local noise events. The nominal scenario label alone is therefore insufficient to determine real-world TSE difficulty; scenario-wise statistics should be interpreted as a coarse diagnostic view rather than a strict ranking of acoustic difficulty.



## License and terms of use

REAL-T is a **public evaluation benchmark**, not a training set.

- **DEV** may be used for hyper-parameter selection, model comparison, and validation. Do **not** train, fine-tune, or run data augmentation on DEV.
- **EVAL1** and **EVAL2** are for final reporting only. Do **not** train, fine-tune, tune hyper-parameters, or otherwise optimize on these splits.
- At test time, process each mix–enroll pair independently. Do not adapt a model using EVAL speaker ids, session names, or other metadata.
- Commercial use is allowed only to the extent permitted by [CC BY-SA 4.0](https://creativecommons.org/licenses/by-sa/4.0/) and the original corpus licenses below.

The REAL-T package (audio clips, transcripts, and overlap annotations) is released under **[CC BY-SA 4.0](https://creativecommons.org/licenses/by-sa/4.0/)**. DEV and EVAL1 are **derived** from the corpora in the table; EVAL2 was recorded for this release. If you redistribute REAL-T or an adapted version of it, you must keep CC BY-SA 4.0 (or a compatible ShareAlike license), give attribution, and cite both REAL-T and every source corpus whose subset you use.


| Subset in REAL-T | Original corpus                                                                   | Original license                                                               | Cite                                     |
| ---------------- | --------------------------------------------------------------------------------- | ------------------------------------------------------------------------------ | ---------------------------------------- |
| AISHELL-4        | [OpenSLR 111](https://www.openslr.org/111/)                                       | [CC BY-SA 4.0](https://creativecommons.org/licenses/by-sa/4.0/)                | Fu et al., Interspeech 2021              |
| AliMeeting       | [OpenSLR 119](https://www.openslr.org/119/)                                       | [CC BY-SA 4.0](https://creativecommons.org/licenses/by-sa/4.0/)                | Yu et al., ICASSP 2022                   |
| AMI              | [AMI Meeting Corpus](https://groups.inf.ed.ac.uk/ami/corpus/)                     | [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)                      | AMI project / Carletta 2007              |
| CHiME-6          | [OpenSLR 150](https://www.openslr.org/150/)                                       | [CC BY-SA 4.0](https://creativecommons.org/licenses/by-sa/4.0/)                | Barker et al. 2018; Watanabe et al. 2020 |
| DiPCo            | [Amazon DiPCo](https://www.amazon.science/publications/dipco-dinner-party-corpus) | [CDLA-Permissive-1.0](https://cdla.io/permissive-1-0/)                         | Van Segbroeck et al., Interspeech 2020   |
| EVAL2            | Collected by the REAL-T organizers                                                | [CC BY-SA 4.0](https://creativecommons.org/licenses/by-sa/4.0/) (this release) | Wang et al., arXiv 2026                  |


CHiME-5/6 were re-issued under CC BY-SA 4.0 as of 1 January 2024 (academic and commercial use; the older Sheffield paid commercial licence no longer applies). AISHELL-4 is CC BY-SA 4.0, not CC BY.

This license covers copyright in the distributed files. It does **not** waive speaker privacy or personality rights. Whether a model trained on ShareAlike speech is “Adapted Material” is jurisdiction-dependent; this README is not legal advice.

Source-corpus citations:

```bibtex
@inproceedings{fu21b_interspeech,
  title     = {{AISHELL-4: An Open Source Dataset for Speech Enhancement, Separation, Recognition and Speaker Diarization in Conference Scenario}},
  author    = {Yihui Fu and Luyao Cheng and Shubo Lv and Yukai Jv and Yuxiang Kong and Zhuo Chen and Yanxin Hu and Lei Xie and Jian Wu and Hui Bu and Xin Xu and Jun Du and Jingdong Chen},
  year      = {2021},
  booktitle = {{Interspeech 2021}},
  pages     = {3665--3669},
  doi       = {10.21437/Interspeech.2021-1397},
  issn      = {2958-1796},
}

@inproceedings{Yu2022M2MeT,
  title     = {M2{M}e{T}: The {ICASSP} 2022 Multi-Channel Multi-Party Meeting Transcription Challenge},
  author    = {Fan Yu and Shiliang Zhang and Yihui Fu and Lei Xie and Siqi Zheng and Zhihao Du and Weilong Huang and Pengcheng Guo and Zhijie Yan and Bin Ma and Xin Xu and Hui Bu},
  booktitle = {{ICASSP} 2022 - 2022 {IEEE} International Conference on Acoustics, Speech and Signal Processing ({ICASSP})},
  year      = {2022},
  pages     = {6167--6171},
  doi       = {10.1109/ICASSP43922.2022.9746465},
  publisher = {{IEEE}},
}

@article{carletta2007ami,
  title   = {Unleashing the killer corpus: experiences in creating the multi-everything {AMI} Meeting Corpus},
  author  = {Jean Carletta},
  journal = {Language Resources and Evaluation},
  volume  = {41},
  number  = {2},
  pages   = {181--190},
  year    = {2007},
  doi     = {10.1007/s10579-007-9040-x},
}

@inproceedings{barker18_interspeech,
  title     = {{The Fifth 'CHiME' Speech Separation and Recognition Challenge: Dataset, Task and Baselines}},
  author    = {Jon Barker and Shinji Watanabe and Emmanuel Vincent and Jan Trmal},
  year      = {2018},
  booktitle = {{Interspeech 2018}},
  pages     = {1561--1565},
  doi       = {10.21437/Interspeech.2018-1768},
  issn      = {2958-1796},
}

@inproceedings{watanabe20b_chime,
  title     = {{CHiME-6 Challenge: Tackling Multispeaker Speech Recognition for Unsegmented Recordings}},
  author    = {Shinji Watanabe and Michael Mandel and Jon Barker and Emmanuel Vincent and Ashish Arora and Xuankai Chang and Sanjeev Khudanpur and Vimal Manohar and Daniel Povey and Desh Raj and David Snyder and Aswin Shanmugam Subramanian and Jan Trmal and Bar Ben Yair and Christoph Boeddeker and Zhaoheng Ni and Yusuke Fujita and Shota Horiguchi and Naoyuki Kanda and Takuya Yoshioka and Neville Ryant},
  year      = {2020},
  booktitle = {{6th International Workshop on Speech Processing in Everyday Environments (CHiME 2020)}},
  pages     = {1--7},
  doi       = {10.21437/CHiME.2020-1},
}

@inproceedings{segbroeck20_interspeech,
  title     = {{DiPCo — Dinner Party Corpus}},
  author    = {Maarten Van Segbroeck and Ahmed Zaid and Ksenia Kutsenko and Cirenia Huerta and Tinh Nguyen and Xuewen Luo and Björn Hoffmeister and Jan Trmal and Maurizio Omologo and Roland Maas},
  year      = {2020},
  booktitle = {{Interspeech 2020}},
  pages     = {434--436},
  doi       = {10.21437/Interspeech.2020-2800},
  issn      = {2958-1796},
}
```



## Citation

If you use REAL-T, please cite:

```bibtex
@inproceedings{li25da_interspeech,
  title     = {{REAL-T: Real Conversational Mixtures for Target Speaker Extraction}},
  author    = {Shaole Li and Shuai Wang and Jiangyu Han and Ke Zhang and Wupeng Wang and Haizhou Li},
  year      = {2025},
  booktitle = {{Interspeech 2025}},
  pages     = {1923--1927},
  doi       = {10.21437/Interspeech.2025-2662},
  issn      = {2958-1796},
}

@misc{wang2026slt2026realtsechallenge,
  title         = {SLT 2026 REAL-TSE Challenge: Real-world Target Speaker Extraction from Conversational Recordings},
  author        = {Shuai Wang and Zihan Qian and Ke Zhang and Jiangyu Han and Zikai Liu and Xiaoyang Yu and Haoyu Li and Marc Delcroix and Kai Yu and Lei Xie and Ming Li and Haizhou Li},
  year          = {2026},
  eprint        = {2607.15198},
  archivePrefix = {arXiv},
  primaryClass  = {eess.AS},
  url           = {https://arxiv.org/abs/2607.15198},
}
```

Contact: [realtse.challenge@gmail.com](mailto:realtse.challenge@gmail.com) (dataset / challenge), [shuaiwang@nju.edu.cn](mailto:shuaiwang@nju.edu.cn) (paper). If you find issues in the audio, labels, transcripts, or documentation, please contact us — reports are welcome and help improve the release.