SecoustiCodec: Cross-Modal Aligned Streaming Speech Codec

Ultra-low bitrate speech codec (0.27-1 kbps) with cross-modal alignment and real-time capabilities

Key Features ✨

• Ultra-Low Bitrate (0.27-1 kbps)
• Cross-Modal Alignment (text-speech synchronization)
• Speaker Preservation (semantic-paralinguistic disentanglement)
• Real-Time Processing (streaming architecture)
• High Efficiency (VAE+FSQ quantization)

Quick Start 🚀

0. System Requirements

# Ubuntu/Debian
sudo apt install sox libsox-dev ffmpeg

# macOS (via Homebrew)
brew install sox ffmpeg

1. Clone & Install:

git clone https://github.com/QiangChunyu/SecoustiCodec.git
cd SecoustiCodec
conda create -n secousticodec python=3.10
conda activate secousticodec
pip install -r requirements.txt

2. Download Models:

pip install huggingface-hub
hf download qiangchunyu/SecoustiCodec --local-dir pretrain_models

3. Run Inference:

# For 86Hz model
python inference.py \
  --source ./test/test.wav \
  --output_dir ./test \
  --model_version 86 \
  --model_path ./pretrain_models/secousticodec_86_hz.pt \
  --hifi_path ./pretrain_models/hift.pt

# For 21.5Hz model
python inference.py \
  --source ./test/test.wav \
  --output_dir ./test \
  --model_version 21.5 \
  --model_path ./pretrain_models/secousticodec_21_5_hz.pt \
  --hifi_path ./pretrain_models/hift.pt

Output Files

File Pattern	Description
{source}_acoustic.wav	Reconstruction from acoustic embeddings
{source}_semantic.wav	Reconstruction from semantic tokens

Model Comparison 🧪

Model Version	Frame Rate	Bitrate
21.5	21.5 Hz	0.27 kbps
86	86 Hz	1.0 kbps

Architecture Overview 🏗️

Citation 📚

@article{qiang2025secousticodec,
  title={SecoustiCodec: Cross-Modal Aligned Streaming Single-Codecbook Speech Codec},
  author={Qiang, Chunyu and Wang, Haoyu and Gong, Cheng and Wang, Tianrui and Fu, Ruibo and Wang, Tao and Chen, Ruilong and Yi, Jiangyan and Wen, Zhengqi and Zhang, Chen and Wang, Longbiao and Dang, Jianwu and Tao, Jianhua},
  journal={arXiv preprint arXiv:2508.02849},
  year={2025}
}

@article{qiang2025vq,
  title={VQ-CTAP: Cross-Modal Fine-Grained Sequence Representation Learning for Speech Processing},
  author={Qiang, Chunyu and Geng, Wang and Zhao, Yi and Fu, Ruibo and Wang, Tao and Gong, Cheng and Wang, Tianrui and Liu, Qiuyu and Yi, Jiangyan and Wen, Zhengqi and Zhang, Chen and Che, Hao and Wang, Longbiao and Dang, Jianwu and Tao, Jianhua},
  journal={IEEE Transactions on Audio, Speech and Language Processing},
  year={2025},
  publisher={IEEE}
}

Acknowledgments 🙏

• HiFiGAN for waveform generation
• MIMICodec for implementation reference

License

SecoustiCodec is released under the Apache License 2.0. See LICENSE for details.