This paper examines the suitability of a recurrent neural network, specifically the long/short term memory (LSTM) cell, for black box modelling of electrodynamic loudspeakers. The goal is to develop a versatile and generic nonlinear model that can be applied in industrial settings, such as distortion cancellation and excursion or power limiters. The presented model has the form of a discrete-time single input multiple output (SIMO) system that takes in a digital audio signal and produces membrane displacement and voice coil current as outputs. The training and validation signals used in the model are described, and data from a two-inch broadband loudspeaker driver is used to train the model. The trained LSTM-based model is then compared to a classical state-space model containing the standard displacement-related nonlinearities of force factor Bl(x), inductance L(x) and compliance Cms(x). The parameters of the state-space model were identified using an industry standard method applied to the same two-inch driver. Results show that the LSTM model outperforms the nonlinear state-space model in both time and frequency domains, although it requires longer training time and has a larger model size. A more detailed model comparison follows, and the results are discussed.
https://www.aes.org/e-lib/browse.cfm?elib=22091
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.
Learn more about the AES E-Library
Start a discussion about this Transducers!
