A refined query-by-example approach to spoken-term-detection on ESL learners' speech

A refined Query-by-Example (QbE) approach is proposed to improve Spoken-Term-Detection (STD) performance on L2 English learners' speech data. A Hidden Markov Model (HMM) is built for each keyword and a computationally efficient, iterative Viterbi decoding is adopted to detect spoken keywords in test. The approach is evaluated on an English as Second Language (ESL) speech database collected over L2 learners with different English proficiency levels. The experimental results show that the new approach achieves a performance better than the traditional DTW-based QbE. Also, it is comparable to that of an LVCSR-based STD but with significant lower complexities and computations. The refined QbE and LVCSR approach to STD are complementary to each other. By fusing the two systems together, we can further improve the MAP and MP@N performance by 6.1%-13.4% and 7.5%-14.4%, respectively, in testing sets of 3 different English proficiency levels over the best performance of either system.