Slyne/funcodec_codecSuperb

Setup Environment

git clone https://github.com/Slyne/FunCodec
cd FunCodec && git checkout slyne_fix && cd ..

The tested environment for the below part is based on docker nvcr.io/nvidia/pytorch:24.04-py3 OR a conda environment should be good as well.

# mount the current directory to /ws; You can put your data in your current
# directory as well.
docker run --gpus all -it -v $PWD:/ws  nvcr.io/nvidia/pytorch:24.04-py3

Or

conda create -n funcodec python=3.10

Install packages

cd /ws/FunCodec;
pip install --editable ./ ; pip install torchaudio; 

Prepare dataset

Please prepare your dataset similar to ${sampling_rate}_wav.scp and put them in /ws/test_wavscp/

44100_wav.scp
48000_wav.scp
16000_wav.scp

Each wav.scp file looks like below:

<wavid> <absolute_path>
WAbHmvQ9zME_00002 /raid/slyne/codec_evaluation/Codec-SUPERB/data/vox1_test_wav/wav/id10302/WAbHmvQ9zME/00002.wav

Example Please follow here to download Codec-SUPERB test datasets.

# suppose the unzip data dir is /ws/data
python3 generate_wavscp.py --input_dir=/ws/data

Download models

Download models from here. And put them under FunCodec/egs/codecSuperb/models

Do inference

Please refer to FunCodec/egs/codecSuperb/do_codecSuperb_infer.sh to do inference.

# set model to the default model trained with 16khz data
model_dir=models/16k/
model_name=8epoch.pth
sample_rates=(16000 44100 48000)  # the input wavscp sample rate ca be 16khz, 44.1khz or 48khz

Run:

cd FunCodec/egs/codecSuperb/
# modify the ref_audio_dir and syn_audio_dir
bash do_codecSuperb_infer.sh