# Bengali Speaker Diarization - Fine-tuned Segmentation Model

Fine-tuned `pyannote/segmentation-3.0` for Bengali speaker diarization.

## Results
- **Best DER**: 0.1312

## Training Data (~370 hours)
- **Synthetic V4**: 600 files (~300 hours) from smam/bengali-diarization-synthetic-v4
- **DISPLACE24**: 67 files (~35 hours)
- **DISPLACE26**: 125 files (~35 hours)

## Training
- Train samples: 600
- Val samples: 67
- Max epochs: 30 (with early stopping on DER, patience=5)
- Optimizer: AdamW (lr=0.0001, weight_decay=0.01)
- Precision: 16-mixed on H100

## Task Configuration
- Duration: 10.0s per chunk
- Max speakers per chunk: 3
- Max speakers per frame: 2 (powerset encoding with overlap support)

## Usage
```python
from pyannote.audio import Model
import torch

# Load weights
state_dict = torch.load('pytorch_model.bin')
model = Model.from_pretrained('pyannote/segmentation-3.0')
model.load_state_dict(state_dict)
```