---

Version update

• 2024.10.12: Important update (please download the latest version of dataset and baseline code)
  • ☑ Dataset updated
    • save the wavform files in dp_speech of train.rar, val.rar and test.rar in 24-bit format to minimize weak background noise (replacing the 16-bit format used in the previous version)
    • correct several inaccurate speaker azimuth annotations, and add annotations for speaker elevation and distance in *_*_source_location.csv
    • update dataset_info.rar
  • ☑ Baseline code updated
    • adjust the speech-recording-to-noise-recording ratio for baseline model training from [0, 15] dB to [-10, 15] dB
  • ☑ Paper updated
    • revise and improve the description of the RealMAN dataset, baseline experiments and other relevant sections
• 2024.06: Inital release

---

Introduction

Motivation: The training of deep learning-based multichannel speech enhancement and source localization systems relies heavily on the simulation of room impulse response and multichannel diffuse noise, due to the lack of large-scale real-recorded datasets. However, the acoustic mismatch between simulated and real-world data could degrade the model performance when applying in real-world scenarios. To bridge this simulation-to-real gap, we presents a new relatively large-scale real-recorded and annotated dataset.

Description: The Audio Signal and Information Processing Lab at Westlake University, in collaboration with AISHELL, has released the Real-recorded and annotated Microphone Array speech&Noise (RealMAN) dataset, which provides annotated multi-channel speech and noise recordings for dynamic speech enhancement and localization:

• Microphone array: A 32-channel microphone array with high-fidelity microphones is used for recording
• Speech source: A loudspeaker is used for playing source speech signals (about 35 hours of Mandarin speech)
• Recording duration and scene: A total of 83.7 hours of speech signals (about 48.3 hours for static speaker and 35.4 hours for moving speaker) are recorded in 32 different scenes, and 144.5 hours of background noise are recorded in 31 different scenes. Both speech and noise recording scenes cover various common indoor, outdoor, semi-outdoor and transportation environments, which enables the training of general-purpose speech enhancement and source localization networks.
• Annotation: To obtain the task-specific annotations, speaker location is annotated with an omni-directional fisheye camera by automatically detecting the loudspeaker. The direct-path signal is set as the target clean speech for speech enhancement, which is obtained by filtering the source speech signal with an estimated direct-path propagation filter.

Baseline demonstration:

• Compared to using simulated data, the proposed dataset is indeed able to train better speech enhancement and source localization networks
• Using various sub-arrays of the proposed 32-channel microphone array can successfully train variable-array networks that can be directly used to unseen arrays

Importance:

• Benchmark speech enhancement and localization algorithms in real scenarios
• Offer a substantial amount of real-world training data for potentially improving the performance of real-world applications

Advantage:

• Realness: Speech and noise are recorded in real environments. Direct recording for moving sources avoids issues associated with the piece-wise generation method. Different individuals move the loudspeaker freely to closely mimic human movements in real applications.
• Quantity and diversity: We record both speech signals and noise signals across various scenes. Compared with existing datasets, our collection offers greater diversity in spatial acoustics (in terms of acoustic scenes, source positions and states, etc) and noise types. This enables effective training of speech enhancement and source localization networks.
• Annotation: We provide detailed annotations for direct-path speech, speech transcriptions and source location, which are essential for accurate training and evaluation.
• Number of channels: The number of microphone channels, i.e. 32, is higher than almost all existing datasets, which facilitates the training of variable-array networks.
• Relatively low recording cost: The recording, playback, and camera devices are portable and easily transportable to different scenes.

Download

To download the entire dataset, you can choose one of the following ways

• directly access the page and download via the download button
  • Hugging Face page
  • AISHELL page
• use the download command:

  huggingface-cli download AISHELL/RealMAN --repo-type dataset --local-dir RealMAN
  

The dataset comprises the following components:

The dataset is organized into the following directory structure:

RealMAN
├── transcriptions.trn
├── dataset_info
│   ├── scene_images
│   ├── scene_info.json
│   └── speaker_info.csv
└── train|val|test|val_raw|test_raw
    ├── train_moving_source_location.csv
    ├── train_static_source_location.csv
    ├── dp_speech
    │   ├── BadmintonCourt2
    │   │   ├── moving
    │   │   │   ├── 0010
    │   │   │   │   ├── TRAIN_M_BAD2_0010_0003.flac
    │   │   │   │   └── ...
    │   │   │   └── ...
    │   │   └── static
    │   └── ...
    ├── ma_speech|ma_noisy_speech
    │   ├── BadmintonCourt2
    │   │   ├── moving
    │   │   │   ├── 0010
    │   │   │   │   ├── TRAIN_M_BAD2_0010_0003_CH0.flac
    │   │   │   │   └── ...
    │   │   │   └── ...
    │   │   ├── static
    │   └── ...
    └── ma_noise

The naming convention is as follows:

# Recorded Signal
[TRAIN|VAL|TEST]_[M|S]_scene_speakerId_utteranceId_channelId.flac

# Direct-Path Signal
[TRAIN|VAL|TEST]_[M|S]_scene_speakerId_utteranceId.flac

# Source Location
[train|val|test]_[moving|static]_source_location.csv

Baseline

• Speech enhancement
• Speech source localization

License

The dataset is licensed under the Creative Commons Attribution 4.0 International (CC-BY-4.0) license.

Citation

To attribute this work, please use the following citation format:

@InProceedings{RealMAN2024,
  author="Bing Yang and Changsheng Quan and Yabo Wang and Pengyu Wang and Yujie Yang and Ying Fang and Nian Shao and Hui Bu and Xin Xu and Xiaofei Li",
  title="RealMAN: A Real-Recorded and Annotated Microphone Array Dataset for Dynamic Speech Enhancement and Localization",
  booktitle="International Conference on Neural Information Processing Systems (NIPS)", 
  year="2024",
  pages=""}