sujalgawas/crnn-vad

CRNN VAD Model

Binary voice activity detection model (SPEECH vs NOISE) trained on Google Speech Commands.

Architecture

• CNN frontend with residual blocks (freq-only pooling to preserve time axis)
• Bidirectional GRU backend for temporal modelling
• Trained with CrossEntropyLoss + label smoothing

Labels

• 0 → NOISE
• 1 → SPEECH

Usage

from huggingface_hub import hf_hub_download
import torch

# Load state dict
weights_path = hf_hub_download(repo_id="sujalgawas/crnn-vad", filename="vad_state_dict.pt")
model = CRNN_VAD()
model.load_state_dict(torch.load(weights_path, map_location="cpu"))
model.eval()

# Or load full Lightning checkpoint
ckpt_path = hf_hub_download(repo_id="sujalgawas/crnn-vad", filename="crnn_vad.ckpt")
model = CRNN_VAD.load_from_checkpoint(ckpt_path)
model.eval()