| --- |
| language: |
| - en |
| - zh |
| - ml |
| - ja |
| - hi |
| - ko |
| - te |
| - ta |
| - pa |
| - fr |
| - bn |
| - kn |
| - ar |
| - th |
| - de |
| - es |
| pretty_name: MISP-M3SD |
| license: apache-2.0 |
| task_categories: |
| - audio-classification |
| tags: |
| - speaker-diarization |
| - audio-visual |
| - multimodal |
| - multilingual |
| - dataset |
| - robust-speech-processing |
| - in-the-wild |
| --- |
| # MISP-M3SD |
|
|
|
|
| ## Dataset Summary |
|
|
| **MISP-M3SD** is a large-scale **multimodal**, **multi-scenario**, and **multilingual** dataset for **robust speaker diarization**, constructed from in-the-wild online videos. It contains more than **770 hours** of synchronised audio-visual recordings, covering **14 scenarios** and **16 languages**. |
|
|
| The dataset is designed to support the development of speaker diarization systems with stronger **cross-domain generalisation** under realistic conditions, including background noise, reverberation, overlapping speech, off-screen speech, motion blur, unstable speaker visibility, and camera switching. |
|
|
| To enable scalable construction, MISP-M3SD is built through a largely automated pipeline, including: |
|
|
| 1. multilingual multi-scenario media acquisition, |
| 2. data cleaning and preprocessing, and |
| 3. cross-modal consistency-guided annotation with selective manual verification. |
|
|
|
|
| ## Access |
|
|
| To use MISP-M3SD: |
|
|
| 1. Download [`audio.zip`](./audio.zip) together with all split files ([`audio.z01`](./audio.z01), [`audio.z02`](./audio.z02), [`audio.z03`](./audio.z03), [`audio.z04`](./audio.z04), [`audio.z05`](./audio.z05), [`audio.z06`](./audio.z06)); |
| 2. Extract the audio archive to obtain the released WAV files; |
| 3. Use [`oracle.rttm`](./oracle.rttm) as the final diarization annotation file; |
| 4. Because of storage limitations, the videos are not distributed as a complete packaged archive. Instead, we provide metadata for each source video, including the corresponding video ID, in [`video_information.xlsx`](./video_information.xlsx), together with video download scripts in the GitHub repository. This allows users to retrieve the videos directly from the original platforms. |
|
|
| ## Key Features |
|
|
| - **Large scale**: 770.55 hours of synchronised audio-visual recordings |
| - **Multilingual**: 16 languages |
| - **Multi-scenario**: 14 scenarios |
| - **Rich interaction complexity**: 7,276 speakers in total, with 5.30 speakers per sample on average |
| - **Rich metadata**: includes source video identifiers, duration, title, description, language, and scenario |
| - **Realistic conditions**: collected from in-the-wild online videos |
| - **Reliable annotation**: cross-modal consistency-guided annotation with selective manual verification |
| - **Practical release format**: audio and annotations are directly provided, while source videos can be retrieved through released scripts and video IDs |
|
|
| ## Scenarios |
|
|
| MISP-M3SD covers 14 scenarios: |
|
|
| - Lesson |
| - Interview |
| - News |
| - Debate |
| - Discussion |
| - Conversation |
| - Job Interview |
| - Meeting |
| - Lecture |
| - Tutorial |
| - Entertainment Vlog |
| - Home Interaction |
| - Dinner Party |
| - Other |
|
|
| The scenario distribution is diverse but naturally uneven, reflecting the characteristics of publicly accessible online videos rather than an artificially balanced design. |
|
|
| ## Data Splits |
|
|
| The dataset is divided into **train / dev / eval** splits at the sample level, with the split assignment for each sample provided in [`split.txt`](./split.txt). |
|
|
| | Split | #Samples | Duration (h) | Avg. Duration (min) | Median Duration (min) | Speech Activity (h) | #Speakers | Avg. Speakers | #Languages | #Scenarios | |
| |---|---:|---:|---:|---:|---:|---:|---:|---:|---:| |
| | Train | 1272 | 716.54 | 33.80 | 18.38 | 656.17 | 6756 | 5.31 | 14 | 14 | |
| | Dev | 50 | 27.10 | 32.52 | 28.42 | 25.71 | 275 | 5.50 | 11 | 13 | |
| | Eval | 50 | 26.91 | 32.30 | 29.62 | 25.04 | 245 | 4.90 | 11 | 13 | |
| | Total | 1372 | 770.55 | 33.70 | 20.98 | 706.91 | 7276 | 5.30 | 16 | 14 | |
|
|
| The splits are constructed to preserve the diversity of the full dataset in terms of scenario, language, duration, speaker-number distribution, and overlap characteristics. |
|
|
|
|
| A comparison with representative audio-visual speaker diarization datasets is shown below. |
|
|
| | Dataset | Scenarios | #Samples | Speakers | Duration (h) | Speech (%) | Noise | Languages | |
| |---|---|---:|---|---:|---:|---|---| |
| | AMI | Meetings | 170 | 3-5 | 100 | 80.91 | No | EN | |
| | AVDIAR | Chat | 27 | 1-4 | 0.35 | 82.6 | No | EN, FR | |
| | AVA-AVD | Daily Activities | 351 | 2-24 | 29.25 | 45.95 | Yes | EN, FR, ZH, DE, KO, ES | |
| | MSDWild | Vlogs | 3143 | 2-10 | 80 | 91.29 | Yes | EN, ZH, TH, KO, JA, DE, PT, AR | |
| | MISP2021&2022 | Conversations | 373 | 2-6 | 121 | 92.30 | Yes | CN | |
| | **MISP-M3SD** | Lesson, Interview, News, Debate, Meeting, Home Interaction, etc. | **1372** | **1-19** | **770.55** | **91.74** | **Yes** | **EN, ZH, ML, JA, HI, KO, TE, TA, PA, etc.** | |
|
|
| Compared with previous datasets, MISP-M3SD provides: |
|
|
| - substantially larger scale |
| - broader scenario coverage |
| - richer multilingual content |
| - more realistic in-the-wild conditions |
| - a scalable construction pipeline for robust multimodal diarization research |
