MADHAV-SAMDANI/speechguard-kws

SpeechGuard KWS — BC-ResNet-8 Keyword Spotter

Part of the SpeechGuard AI system submitted to Samsung EnnovateX AX Hackathon 2026.

Model Description

BC-ResNet-8 keyword spotter trained on Google Speech Commands v2 with noise augmentation. Uses PCEN (Per-Channel Energy Normalization) frontend for robust noise handling.

Performance

Metric	Value
TA Clean	99.0%
TA Noisy (-5 to +30 dB)	98.5%
Parameters	2,444
Latency (CPU)	1.1ms

Usage

import torch
from huggingface_hub import hf_hub_download

# Download checkpoint
ckpt_path = hf_hub_download(
    repo_id="MADHAV-SAMDANI/speechguard-kws",
    filename="best_kws.pt"
)

# Load model
from speechguard.kws.bc_resnet import BCResNet8
ckpt  = torch.load(ckpt_path, map_location="cpu", weights_only=False)
model = BCResNet8(num_classes=len(ckpt["classes"]), n_mels=80)
model.load_state_dict(ckpt["model_state"])
model.eval()

Training

• Dataset: Google Speech Commands v2 (2000 samples/class)
• Epochs: 35
• Optimizer: AdamW with cosine LR annealing
• Noise augmentation: ESC-50 + synthetic (white, pink, babble)
• Hardware: MacBook Air CPU (~70 minutes)

Citation

Samsung EnnovateX AX Hackathon 2026 — Problem #04 Team: Placecomm Prophets (IIT Kharagpur)