Metadata-Version: 2.1
Name: jadia
Version: 0.1.1
Summary: JaNet diarization package
Author-email: Sergey Skrebnev <sergey.skrebnev@gmail.com>
Project-URL: Homepage, https://github.com/skrbnv/jadia
Classifier: Programming Language :: Python :: 3
Classifier: License :: OSI Approved :: MIT License
Classifier: Operating System :: OS Independent
Requires-Python: >=3.0
Description-Content-Type: text/markdown
License-File: LICENSE
Requires-Dist: librosa>=0.10.1
Requires-Dist: numpy>=2.1.0
Requires-Dist: silero_vad>=5.1
Requires-Dist: soundfile>=0.12.1
Requires-Dist: soxr>=0.3.7
Requires-Dist: torch>=2.4.0

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