A PyTorch port of VGGish1,
a feature embedding frontend for audio classification models. The weights are ported directly from the tensorflow model, so embeddings created using torchvggish
will be identical.
There are two options: you can install the last stable version from pypi, or clone this repo and install.
# optional: create virtual env
cd torchvggish && python3 -m venv .env
source activate .env/bin/activate
pip install -i https://test.pypi.org/simple/ torchvggish==0.1
# OR get the latest version
git clone [email protected]:harritaylor/torchvggish.git
pip install -r requirements.txt
Barebones example of creating embeddings from an example
wav file:
from torchvggish import vggish, vggish_input
# Initialise model and download weights
embedding_model = vggish()
embedding_model.eval()
example = vggish_input.wavfile_to_examples("example.wav")
embeddings = embedding_model.forward(example)
1. S. Hershey et al., ‘CNN Architectures for Large-Scale Audio Classification’,\ in International Conference on Acoustics, Speech and Signal Processing (ICASSP),2017\ Available: https://arxiv.org/abs/1609.09430, https://ai.google/research/pubs/pub45611