Efficient and Robust Music Identification With Weighted Finite-State Transducers

Abstract

Abstract-We present an approach to music identification based on weighted finite-state transducers and Gaussian mixture models, inspired by techniques used in large-vocabulary speech recognition. Our modeling approach is based on learning a set of elementary music sounds in a fully unsupervised manner. While the space of possible music sound sequences is very large, our method enables the construction of a compact and efficient representation for the song collection using finite-state transducers.This paper gives a novel and substantially faster algorithm for the construction of factor transducers, the key representation of song snippets supporting our music identification technique. The complexity of our algorithm is linear with respect to the size of the suffix automaton constructed. Our experiments further show that it helps speed up the construction of the weighted suffix automaton in our task by a factor of 17 with respect to our previous method using the intermediate steps of determinization and minimization. We show that, using these techniques, a large-scale music identification system can be constructed for a database of over 15 000 songs while achieving an identification accuracy of 99.4% on undistorted test data, and performing robustly in the presence of noise and distortions.

Mobile music modeling, analysis and recognition

“…Accuracy for multi-conditional training Front-end. We test the modifications we propose in Section 2.1 to our baseline speech recognition front-end (similar to that reported in [1]) by training an acoustic model on CLEAN, building the FST from the songs in INDEX, and testing the modifications on the four sets of snippets. As shown in Table 2 adapting the front-end to the task of music recognition consistently improved identification accuracy across the different test sets.…”

“…The above system achieved 99.9% identification accuracy over test snippets cut from clean recordings, and in [6,1] we showed that the system was robust to synthetic distortions. Nevertheless, we observed a significant degradation in accuracy when the test recordings were recorded with mobile phones.…”

“…The use of weighted finite-state transducers (WFSTs) for music recognition was introduced in [2,1]. In this approach, the acoustic model is initialized by unsupervised clustering of the features to form a set of music phonemes.…”

“…The algorithm for constructing this WFST uses weighted determinization and minimization, which preserves the total weight of a path corresponding to a given song. We refer the reader to [1] for details.…”

“…

ABSTRACTWe present an analysis of music modeling and recognition techniques in the context of mobile music matching, substantially improving on the techniques presented in [1]. We accomplish this by adapting the features specifically to this task, and by introducing new modeling techniques that enable using a corpus of noisy and channel-distorted data to improve mobile music recognition quality.

…”