data_preparation.md

Data preparation

Format description

The training set should be arranged as follows:

PATH_AUDIO_FILES  
│
└───spkr1
│   └───...
│         │   seq_11.wav
│         │   seq_12.wav
│         │   ...
│   
└───spkr2
    └───...
          │   seq_21.wav
          │   seq_22.wav

The spkrN information can then be used to sample sequences when training the model (within- or across- sampling)

Pre-filtering strategies

1. Remove non-speech segments

If you want to discard non-speech segments from your training set, you can use:

• The voice type classifier if your dataset consists of long-form recordings
• However, a pretrained pyannote.audio voice activity detection model will likely return more accurate segments if your data is clean

Both return .rttm files indicating where there is speech in each audio file. If you used the voice type classifier (as I did) you can run:

python data/extract_segments.py --audio_path /path/to/audio --rttm_path /path/to/rttm --output_path /path/to/output \
  --durations [128] --sampling random --classes [MAL,FEM] --min_dur 1.5

This will create a training set of 128h of audio with segments produced by male and female speakers whose duration is at least 1.5s. Then, you can run:

python data/build_power_two_training.py --audio_path /path/to/speech/segments --output_path=/path/to/output \
  --nb_packets=16 --duration=28800

This will create 16 training sets of 8 hours (28800 seconds), merged two by two to create 8 training sets of 16 hours, ... until a single training set of 128 hours is obtained.

2. Remove noisy and/or reverberated segments

If you want to discard noisy and/or reverberated segments from your training set, you can use Brouhaha.