Published Paper Details:

This paper presents a complete end-to-end text-based Spoken Term Detection (STD) system evaluated on the QUESST 2014 multilingual benchmark, covering six low-resource European languages: Albanian, Basque, Czech, non-native English (NNEnglish), Romanian, and Slovak. The proposed system employs the HappyScribe cloud-based Automatic Speech Recognition (ASR) service for audio transcription, followed by the Terrier Information Retrieval (IR) framework with the Hiemstra language model (Hiemstra_LM) for inverted index construction and probabilistic ranked retrieval. Detection performance is assessed using the official QUESST 2014 metrics: minimum Normalised Cross-Entropy (minCNXE, lower is better) and Maximum Term-Weighted Value (MTWV, higher is better), across three query types: Type 1 (exact match), Type 2 (lexical variation), and Type 3 (multi-word expression). The text-based system is systematically compared against four speech-to-speech acoustic matching baselines: Dynamic Time Warping (DTW), Convolutional Neural Network (CNN) matching, End-to-End neural matching, and the Water Chicken Swarm Optimisation-based Deep Segmental Neural Network (WCSO-DSNN). The text-based system achieves a best multilingual minCNXE of 0.2489 and MTWV of 0.8337 for Type 1 evaluation queries, substantially outperforming all four acoustic baselines. Language-specific analysis establishes Romanian (minCNXE=0.1012, MTWV=0.9123) and Slovak (minCNXE=0.1115, MTWV=0.9123) as the highest-performing languages, while NNEnglish (best minCNXE=0.3983) is the most challenging due to multi-speaker accent variability. The findings demonstrate that text-based STD is significantly more mature than speech-to-speech acoustic matching for settings where reliable ASR is available, while acoustic methods retain a practical advantage in true zero-resource environments.