H-GTCRN

This repository is the official implementation of the Interspeech2025 paper: A Lightweight Hybrid Dual Channel Speech Enhancement System under Low-SNR Conditions. For more details, please refer to the ISCA Archive.

Figure 1: The framework of our proposed system.

• [2026-3-13] The model implementation and pre-trained checkpoint are released.
• [2025-8-17] The paper is uploaded to ISCA Archive.
• [2025-5-25] The paper is uploaded to arxiv .

To run inference on audio files, use:

python infer.py --input_dir <input_dir> --output_dir <output_dir> --checkpoint <checkpoint> --device <device> --suffix <suffix>

The directory structure of the audio samples is shown below:

    samples
    ├── Samples1
    |   ├── Samples1_clean.wav
    |   ├── Samples1_noisy.wav
    |   ├── Samples1_IVA.wav
    |   ├── Samples1_GTCRN.wav
    |   ├── Samples1_DC_GTCRN.wav
    |   └── Samples1_Proposed.wav
    | ...
    └── Samples3
        ├── Samples3_clean.wav
        ├── Samples3_noisy.wav
        ├── Samples3_IVA.wav
        ├── Samples3_GTCRN.wav
        ├── Samples3_DC_GTCRN.wav
        └── Samples3_Proposed.wav

If you find this work useful, please cite our paper:

@inproceedings{wang2025lightweight,
  title={A Lightweight Hybrid Dual Channel Speech Enhancement System under Low-SNR Conditions},
  author={Wang, Zheng and Rong, Xiaobin and Sun, Yu and Sun, Tianchi and Lin, Zhibin and Lu, Jing},
  booktitle={Proc. Interspeech 2025},
  pages={1178--1182},
  year={2025}
}

We gratefully acknowledge the following resources that made this project possible:

• GTCRN: SOTA lightweight speech enhancement model architecture.
• SE-train: Excellent training code template for DNN-based speech enhancement.
• pyroomacoustics