This was a collage project.
We extract features for every record and create dataset in the schema
audio-based features --> label M or F.
No Deep Learning.
The term gender recognition is unfortunately inaccurate. Every speaker in Librispeech
dataset has M or F label assigned. M voice is considered to have male
characteristics and F voice has female characteristics.
We follow this notation focusing only on technical aspects of classification.
The voice timbre may be very misleading and its subjective
perception is not the area of focus of this project.
Simple deps:
pip install -r requirements.txt
Download and unpack Librispeech data directly in data/raw/ folder so that the data with SPEAKERS.txt, BOOKS.txt etc. files is in
data/raw/LibriSpeech/.
Project is based only on clean files (train-100, train-360, dev and test).
Run:
python extract_features
Trained on signals from train-clean-100 set, tested on signals from dev-clean.
Dev f1 score: 0.8861251457442676