jadia 0.1

JaNet diarization package

Jadia Diarization package

Kmeans-based fully open source (model code provided) package, good at parsing dialogs. Requires number of speakers to be provided. Fast in 'fast_fit' mode, when clusters are defined using only first slice of audio.

Install:

pip install jadia

pip install jadia-plot if you want to plot predictions

Usage

diarizer = Jadia(device=torch.device("cuda:0"),model="lite", batch_size=64)
segments = diarizer.process(FILENAME, num_voices=NUM_VOICES)

or

diarizer = Jadia(device=torch.device("cuda:0"), model="lite", batch_size=64)
diarizer.setup(num_voices=NUM_VOICES)
diarizer.load_audio(filename=FILENAME)
predictions = diarizer.predict()
segments = diarizer.predictions_to_segments(predictions)

Look into eval.ipynb notebook for plotting, metrics etc.

TODO:

• improved lite model
• heavier model with extra transformer layers
• fine-tuning