This research proposes an approach for computing the time offsets between audio sequences that contain musical sounds from different instruments produced in a distributed way and which have a set of weak features that are not useful as alignment points. It is therefore necessary to apply transformations in order to find a set of distinctive features to compute the offset values in a suitable way. The main issue that occurs with such a system is nonlinearity that does not allow the delay to be predicted by using a linear function. To solve this problem, the authors propose a set of long short-term memory (LSTM) layers to create a neural network model capable of learning such features transformations in a supervised approach, using a gradient-descent optimizer. This demonstrates the use of a recurrence matrix to extract timing information from a set of transformed features given by the neural network output. With this approach, the algorithm can classify up to 60% of a specific combination from the MedleyDB data set, and reduce the search space to five possibilities with accuracy up to 90% while keeping the precision of 10 ms. This performance is equal or better than state-of-the-art methods.
Click to purchase paper as a non-member or login as an AES member. If your company or school subscribes to the E-Library then switch to the institutional version. If you are not an AES member and would like to subscribe to the E-Library then Join the AES!
This paper costs $33 for non-members and is free for AES members and E-Library subscribers.