This repository contains files related to my second qualifying paper on voice gender classification using deep neural networks with raw audio as inputs.
This project uses data from the Voices of California corpus. fetch.py handles downloading relevant audio and transcripts from the VOC master directory from the mounted Box drive. speakers.py handles transcripts of many hundreds of sociolinguistic interviews to determine when in the matching recording the interviewee is producing non-overlapping speech.
Model predictions are calculated and stored in anonymized .csv files by process.py.
This project was written in Python 2.7.16, and there are tricky interleaving dependencies, particularly with regard to the pysoundfile module. I can't guarantee it works on any other version.
I owe a great many thanks to Oscar Knagg, who made the code
from his very successful raw-audio voice gender classification experiment public in 2018.
data.py, utils.py and models.py remain essentially unchanged. I have modified process.py to suit my particular needs.
I trained the network myself (on the same LibriSpeech dataset) and achieved very slightly higher accuracy than Knagg reports in his Medium article (99.1% after 7 epochs).