README.md 772 Bytes
Newer Older
Alessio Brutti's avatar
Alessio Brutti committed
1
2
# vggvox_features

Alessio Brutti's avatar
Alessio Brutti committed
3
4
5
6
7
8
9
This code is based on the VGGvox model distributed by Oxford in:
https://github.com/a-nagrani/VGGVox

Models are trained on VoxCeleb2 and are available in model/weights.h5

It finds all the wav files in a given folder (hard-coded) are producess a pickle dataset with a list of arrays containing the VGGvox embeddings for each file.

10
11
12
13
14
15
16
17
18
Usage:

python src/extract_features.py -h 

--input_dir: folder containing the wav files to process. It assumes that files are grouped by session or speaker.
--output_dir: folder where features are stored in a pickle file (one for each session/speaker)
--model: file with model weights.


Alessio Brutti's avatar
Alessio Brutti committed
19
20
21

This code is derived from the python version of VGGvox available here:
https://github.com/linhdvu14/vggvox-speaker-identification