A model for keyword spotting in voice signal for specialized computer systems

Abstract

Keyword spotting in voice signal is a crucial task for specialized, low-resource computer systems, such as ground drones, particularly when operating under challenging conditions with limited computational power and without reliable cloud access. This paper presents a novel, modular model for efficient keyword spotting that does not rely on deep neural networks. The model's core principle is the differential weighting of Mel-Frequency Cepstral Coefficients , prioritizing those coefficients most discriminative for phonetic content. The architecture incorporates robust signal conditioning, dynamic feature extraction (including delta and delta-delta derivatives), the transformation of acoustic features into compact string-based "fingerprints", and final classification using the Levenshtein distance. Experimental validation, conducted on a Ukrainian-language corpus of drone commands with lexicons of up to 200 words, demonstrated the model's high performance and scalability. The system achieved an F1-score of 0.92 under ideal conditions and showed significant resilience in noisy environments, maintaining an F1-score of 0.78 at a 5dB signal-to-noise ratio. Furthermore, the proposed system significantly outperformed a baseline version (using only basic Mel-Frequency Cepstral Coefficients without derivatives or normalization) by up to 33 percentage points in F1-score under challenging conditions. The study validates that this optimized classical Keyword Spotting approach provides an effective and fully autonomous solution for edge computing applications where resource efficiency and independence from cloud infrastructure are paramount, especially in critical scenarios like military operations.