audiosep

AudioSep: Separate Anything You Describe

Input

• Mixed audio file

Audio file in wav format with mixed sources. input.wav

input.mov

This audio file was adapted from the official audiosep implementation

https://audio-agi.github.io/Separate-Anything-You-Describe/demos/exp31_water/drops_mixture.wav

• Text condition

Text description of the sound source you want to separate.

Output

• Audio file

Separated audio source according to the text query.

Saves to ./output.wav by default but it can be specified with the --path option

Usage

Internet connection is required when running the script for the first time, as the model files will be automatically downloaded.

Running this script will separate sound sources from the original input audio file, according to the language query.

Example1: Extract sound of thunder

$ python3 audiosep.py -p "thunder" -i input.wav -s output_thunder.wav

output_thunder.mov

Example2: Extract sound of waterdrops

$ python3 audiosep.py -p "water drops" -i input.wav -s output_waterdrops.wav

output_waterdrops.mov

.wav file containing the sound source separated from the original mixture will be created in both cases.

Reference

• AudioSep
• Separate Anything You Describe

Framework

Pytorch

Model Format

ONNX opset=11

Netron

• audiosep_text.onnx.prototxt
• audiosep_resunet.onnx.prototxt