Robust and Fast Lyric Search based on Phonetic Confusion Matrix

This paper proposes a robust and fast lyric search method for music information retrieval. Current lyric search systems by normal text retrieval techniques are severely deteriorated in the case that the queries of lyric phrases contain incorrect parts due to mishearing and misremembering. To solve this problem, the authors apply acoustic distance, which is computed based on a confusion matrix of an ASR experiment, into DP-based phonetic string matching. The experimental results show that the search accuracy is increased by more than 40% compared with the normal text retrieval method; and by 2% ∼4% compared with the conventional phonetic string matching method. Considering the high computation complexity of DP matching, the authors propose a novel two-pass search strategy to shorten the processing time. By pre-selecting the probable candidates by a rapid index-based search for the first pass and executing a DP-based search among these candidates during the second pass, the proposed method reduces processing time by 85.8% and keeps search accuracy at the same level as that of a complete search by DP matching with all lyrics.