smam/pyannote-segmentation-bengali-multilingual-v2

Bengali Speaker Diarization - Segmentation Model V2

Fine-tuned pyannote/segmentation-3.0 for Bengali speaker diarization (DLSPRINT26 hackathon).

Training Details

• Base Model: pyannote/segmentation-3.0
• Best Validation Accuracy: 78.04%
• Best Epoch: 36/50
• Training Samples: 510
• Validation Samples: 90

Improvements in V2

• Dropout: 30% on output layer
• Weight Decay: 0.01 (L2 regularization)
• Gradient Clipping: 1.0
• Label Smoothing: 0.1
• Learning Rate: OneCycleLR (max=0.0003, 10% warmup)
• Batch Size: 64

Augmentation

• Training: Diverse on-the-fly augmentation (noise, reverb, volume, time masking)
• Validation: Fixed one-time augmentation for consistent evaluation

Usage

from pyannote.audio import Model
import torch

# Load base model
model = Model.from_pretrained("pyannote/segmentation-3.0")

# Load fine-tuned weights
weights = torch.hub.load_state_dict_from_url(
    "https://huggingface.co/smam/pyannote-segmentation-bengali-multilingual-v2/resolve/main/pytorch_model.bin",
    map_location="cpu"
)
model.load_state_dict(weights)