The purpose of this assignment is to attempt to perform audio diarization automatically in order to assist researchers in reducing overall costs and speed up data collection.
This file will do most prep work for the neural net. It will split all the audio up into individual tracks (trackA, trackB). It will then pick out segments with audio, and compair with it's companion track to determin if the audio is clean or dirty, clean if it does not conflict with the other track or dirty if it is. This analysis is cached in order to speed up future runs, but initial run could take hours. The result of this is a list of dictionaries. Each dictionary element has {'original':"string", 'trackA':"string", 'trackB':"string", 'aClassification':[stMilisecond, endMilisecond, "classification"], 'bClassification':[stMilisecond, endMilisecond, "classification"]}
- No Data should be commited!!!! (keep in local machine to help I have created a directory that ignores all .wav files)
You need to install FFMPEG, google. You will also need to pip install pydub. Some of the prep does require a little bit of a specific structure in the data folder, so there is a buildDatafolder.py script to help build this (note it is cross platformish) this should only need to be run once. Put the audio in the data/rawAudio and csv in data/csvFinal and the code will do the rest. After the first run, prepAudio will pull from it's cached files in order to reduce the time of loading preped data to a few minutes, but note this is cpu/ram intensive.