SimWhisper_Codec

<div align="center">

🎙️ SimWhisper-Codec

Speaking Clearly: A Simplified Whisper-Based Codec for Low-Bitrate Speech Coding

<p> <a href="https://zhangxinwhut.github.io/SimWhisper-Codec/"><img src="https://img.shields.io/badge/🎧_Demo-Online-brightgreen" alt="Demo"></a> <a href="https://arxiv.org/pdf/2510.20504v2"><img src="https://img.shields.io/badge/Paper-Arxiv-red" alt="paper"></a> <a href="https://huggingface.co/xxx123456/SimWhisper_Codec"><img src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Model%20Page-yellow" alt="Hugging Face"></a> <a href="https://github.com/ZhangXinWhut/SimWhisper-Codec"><img src="https://img.shields.io/badge/GitHub-Code-black?logo=github" alt="GitHub"></a> </p>

A semantic-first speech codec that achieves superior performance through architectural simplification rather than complex supervision.

</div>

---

✨ Highlights

• 🚀 low Bitrate: Only 1.1 kbps at 16 kHz sampling rate
• 🔊 High Quality Speech Reconstruction: Achieving UTMOS 4.00 WER 2.75 (hubert-large-ls960-ft) sim 0.83 (wavlm_large_finetune) stoi 0.93 pesq-nb 3.29 pesq-wb 2.72 on librispeech-test-clean reconstruction (gt: WER 2.16 UTMOS 4.09)
• 🧊 Frozen Encoder: No fine-tuning of Whisper encoder required
• ⚡ Simple & Efficient: Architectural simplification over complex supervision

📊 Performance

Evaluated on LibriSpeech test-clean

🚀 Quick Start

Installation

# Clone repository
git clone https://github.com/ZhangXinWhut/SimWhisper-Codec.git && cd SimWhisper-Codec

# Create and activate conda environment
conda create -n swcodec python=3.10 -y && conda activate swcodec

# Install dependencies
pip install -r requirements.txt

Available Models 🗂️

Download Model Weights

You need to download the SimWhisper-Codec model weights. You can find the weights in the SimWhisper-Codec Hugging Face repository.

mkdir -p ./weights && huggingface-cli download xxx123456/SimWhisper_Codec SimWhisperCodec.pt --local-dir ./weights/

Inference

python inference.py --input_dir /path/to/LibriSpeech/test-clean

The reconstructed audio files will be available in the output_wavs/ directory.

🙏 Acknowledgements

Our codebase builds upon the XY-Tokenizer. We thank the authors for their excellent work.

📝 Citation

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

@misc{zhang2025speakingclearlysimplifiedwhisperbased,
      title={Speaking Clearly: A Simplified Whisper-Based Codec for Low-Bitrate Speech Coding}, 
      author={Xin Zhang and Lin Li and Xiangni Lu and Jianquan Liu and Kong Aik Lee},
      year={2025},
      eprint={2510.20504},
      archivePrefix={arXiv},
      primaryClass={cs.SD},
      url={https://arxiv.org/abs/2510.20504}, 
}