Skip to content

Latest commit

 

History

History
72 lines (54 loc) · 3.2 KB

README.md

File metadata and controls

72 lines (54 loc) · 3.2 KB
Raw

SuperCodec: A Neural Speech Codec with Selective Back-Projection Network

githubio arXiv GitHub Repo stars GitHub

Updates

  • Code release. (Jul. 27, 2024)
  • Online demo at Github See here. (Aug. 13, 2023)
  • Supports 16-48 kHz at variable bitrates. (Jul. 27, 2024)

In this paper, we present SuperCodec, a neural speech codec that replaces the standard feedforward up- and downsampling layers with Selective Up-sampling Back Projection (SUBP) and Selective Down-sampling Back Projection (SDBP) modules. Our proposed method efficiently preserves the information, on the one hand, and attains rich features from lower to higher layers of the network, on the other. Additionally, we propose a selective feature fusion block in the SUBP and SDBP to consolidate the input feature maps

inference
Supercodec

Pre-requisites

  1. Clone this repo: git clone https://github.com/exercise-book-yq/Supercodec.git

  2. CD into this repo: cd Supercodec

  3. Install python requirements: pip install -r requirements.txt

Training Example

# train
python train.py --config config_v1.json

Inference Example

# inference
python inferece.py --checkpoint_file [generator checkpoint file path]

Additional Experiments

Objective evaluation testing on our test set from VCTK at 16 kHz sampling rate. We compare our proposed method with existing various codecs trained with the same configuration.

Model Bitrate ViSQOL STOI(%) WARP-Q(↓)
Supercodec 1 kbps 3.118 84.80 2.219
TiCodec 1 kbps 2.490 80.21 2.578
HiFiCodec 1 kbps 2.060 75.19 2.840
EnCodec 1 kbps 2.202 76.53 2.687

Objective evaluation testing on our test set from VCTK at 24 kHz sampling rate. We compare our proposed method with existing various codecs trained with the same configuration.

Model Bitrate ViSQOL STOI(%) WARP-Q(↓)
Supercodec 1.5 kbps 3.322 85.61 2.147
TiCodec 1.5 kbps 2.639 79.03 2.539
HiFiCodec 1.5 kbps 2.026 76.80 2.761
EnCodec 1.5 kbps 2.202 79.81 2.569

All models are non-causal and trained on LibriTTS.

References