Comparative Analysis of Post-Training Quantization vs. Quantization-Aware Training for Keyword Spotting: 2025-2026 Advances in Hardware Co-Design

Read full text Summarise with VIRAT
Authors: Amita Contractor, Vaibhavi Pandya
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

The deployment of Deep Learning models on resource-constrained microcontrollers requires significant reduction in model size and computational complexity. This paper presents an updated comparative analysis of Post-Training Quantization (PTQ) and Quantization-Aware Training (QAT) applied to Keyword Spotting (KWS) tasks, incorporating critical advances from 2025-2026 research. Using the Google Speech Commands dataset, we evaluate the performance of Depthwise Separable CNNs (DS-CNN) and BC-ResNet architectures under both quantization paradigms. Recent findings demonstrate that while PTQ remains the dominant fast-deployment route—especially when paired with hardware zero-skipping techniques—fixed-point QAT (FXP-QAT) now enables sub-8-bit models (down to 4- bit) with significant latency and accuracy gains. Notably, 8-bit FXP-QAT models achieve 4–6% improvement in relative false discovery rate at fixed false reject rate compared to full-precision models, together with a 68% reduction in execution time. We further analyze the emerging trend of hardwarealgorithm co-design, where architectures like BCResNet are optimized for ultra-low-power accelerators with hybrid time-feature-frequency-domain zeroskipping, achieving energy consumption as low as 36 nJ per decision. This work synthesizes 98 papers from 2025-2026 to provide a comprehensive view of the current state-of-the-art in quantized KWS for edge deployment.

Fields of Study & Subjects

Advanced Neural Network Applications Speech Recognition and Synthesis Low-power high-performance VLSI design

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
1 January 2026
Language
EN