Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

EdgeSpot: Efficient and High-Performance Few-Shot Model for Keyword Spotting

Oguzhan Buyuksolak, Alican Gok, Osman Erman Okman

TL;DR

EdgeSpot tackles the challenge of few-shot keyword spotting on edge devices by integrating a compact BC-ResNet backbone with a trainable PCEN frontend, early fusion of initial blocks, temporal relative encoding, and a lightweight temporal self-attention head. It uses a teacher–student distillation framework from a self-supervised Wav2Vec2 teacher, combined with Sub-center ArcFace, to produce robust embeddings suitable for prototype-based inference. In MSWC and cross-domain GSC evaluations, EdgeSpot-4 delivers substantial accuracy gains at low false-alarm rates while maintaining a small compute footprint (29.4M MACs) and modest parameter count (128k), closely tracking the SSL teacher in many settings. The results demonstrate strong edge-device practicality and improved cross-domain generalization for FS-KWS without sacrificing efficiency.

Abstract

We introduce an efficient few-shot keyword spotting model for edge devices, EdgeSpot, that pairs an optimized version of a BC-ResNet-based acoustic backbone with a trainable Per-Channel Energy Normalization frontend and lightweight temporal self-attention. Knowledge distillation is utilized during training by employing a self-supervised teacher model, optimized with Sub-center ArcFace loss. This study demonstrates that the EdgeSpot model consistently provides better accuracy at a fixed false-alarm rate (FAR) than strong BC-ResNet baselines. The largest variant, EdgeSpot-4, improves the 10-shot accuracy at 1% FAR from 73.7% to 82.0%, which requires only 29.4M MACs with 128k parameters.

EdgeSpot: Efficient and High-Performance Few-Shot Model for Keyword Spotting

TL;DR

EdgeSpot tackles the challenge of few-shot keyword spotting on edge devices by integrating a compact BC-ResNet backbone with a trainable PCEN frontend, early fusion of initial blocks, temporal relative encoding, and a lightweight temporal self-attention head. It uses a teacher–student distillation framework from a self-supervised Wav2Vec2 teacher, combined with Sub-center ArcFace, to produce robust embeddings suitable for prototype-based inference. In MSWC and cross-domain GSC evaluations, EdgeSpot-4 delivers substantial accuracy gains at low false-alarm rates while maintaining a small compute footprint (29.4M MACs) and modest parameter count (128k), closely tracking the SSL teacher in many settings. The results demonstrate strong edge-device practicality and improved cross-domain generalization for FS-KWS without sacrificing efficiency.

Abstract

We introduce an efficient few-shot keyword spotting model for edge devices, EdgeSpot, that pairs an optimized version of a BC-ResNet-based acoustic backbone with a trainable Per-Channel Energy Normalization frontend and lightweight temporal self-attention. Knowledge distillation is utilized during training by employing a self-supervised teacher model, optimized with Sub-center ArcFace loss. This study demonstrates that the EdgeSpot model consistently provides better accuracy at a fixed false-alarm rate (FAR) than strong BC-ResNet baselines. The largest variant, EdgeSpot-4, improves the 10-shot accuracy at 1% FAR from 73.7% to 82.0%, which requires only 29.4M MACs with 128k parameters.
Paper Structure (10 sections, 9 equations, 2 figures, 1 table)

This paper contains 10 sections, 9 equations, 2 figures, 1 table.

Table of Contents

  1. Introduction
  2. PROPOSED METHOD
  3. Model Architecture
  4. Per-Channel Energy Normalization (PCEN)
  5. Fusion of Early Blocks
  6. Relative Positional Encoding
  7. Scaled Dot Product Attention (SDPA)
  8. Training Methodology
  9. Experimental Results
  10. Discussion

Figures (2)

  • Figure 1: Architecture of the Proposed Model. The model processes a 40×101 Mel-Spectrogram input and applies Per-Channel Energy Normalization (PCEN) for preprocessing. It then passes through a 5×5 Conv2D layer followed by depthwise 2D convolutions (DW Conv2D) and residual blocks (BC-Res. and Fused BC-Res.) to extract hierarchical features. Positional encoding is incorporated using a depthwise Conv1D layer to enhance temporal feature representation. A self-attention mechanism (SDPA) focuses on capturing dependencies in the time dimension. Finally, a Conv1D layer reduces dimensionality, producing a 64-dimensional embedding as the output.
  • Figure 2: Comparison of (a) BC-ResBlock and (b) Fused BC-ResBlock, highlighting the replacement of 'Temporal DW Conv. + 1x1 Conv.' with a single 'Temporal Conv.'.