DOOM-SED

Downstream-oriented open-vocabulary model for sound event detection

Usage

Installation

Install M3L:

pip install m3l-toolkit

BEATs checkpoint

Download the pre-trained BEATs_iter3_plus_AS2M.pt checkpoint from the official BEATs repository.

Set the BEATS_CKPT_PATH environment variable to the directory containing the checkpoint file, not the file itself:

export BEATS_CKPT_PATH=/path/to/beats

The checkpoint must be located at:

$BEATS_CKPT_PATH/BEATs_iter3_plus_AS2M.pt

This checkpoint is required for inference and is not included in this model repository.

Inference

Create a label_names.json file containing a JSON array of sound event labels to detect, for example:

["Dog barking", "Siren", "Footsteps"]

Run inference with:

python -m m3l.audiotext.ov_sed.pipelines.detect one \
    --disable_strict=false
    hf://b-sigpro/doom-sed \
    label_names.json \
    input.wav \
    output.pkl \

Input audio is converted to mono at 16 kHz. The output is a pickle file containing prediction scores (y_pred) and the corresponding label names (label_names).

Citation

If you use this model in your research, please cite the following work:

@inproceedings{sakurai2026doomsed,
  title = {{DOOM-SED}: Downstream-Oriented Open-Vocabulary Model for Sound Event Detection},
  author = {Shun Sakurai and Yudai Kandori and Yoshiaki Bando and Keisuke Imoto and Masaki Onishi},
  booktitle = {Proceedings of the 11th Workshop on Detection and Classification of Acoustic Scenes and Events (DCASE 2026)},
  pages = {245--249},
  year = {2026}
}

License

This model is released under the apache-2.0 license.

Note that this model is built on the following base models and is also subject to their respective license terms:

• LAION-CLAP — apache-2.0

• BEATs — mit