AES 5th International Conference on Automotive Audio

Analysing and simulating loudspeaker systems in vehicles is a complex task. In order to be able to make a statement about the sound characteristics at an early stage of development, various physical disciplines must be coupled together. The simulation starts with electromagnetism and is followed by mechanical modelling, which in turn is linked to the acoustic model via the diaphragm speed. Some audio systems consist of more than 20 loudspeakers that are used simultaneously, each of which is designed for a different frequency range. A rough categorisation is made here into the low, mid and high frequency ranges, whereby each of the ranges presents different difficulties for the simulation.
The focus of this article is on the high frequency range, which places great demands on the acoustic simulation with classical discretisation methods, such as the finite element method (FEM) or the boundary element method (BEM), particularly due to the short wavelengths. In order to obtain the prediction in the entire frequency range up to 20,000 Hz, either accelerated methods, such as the fast multipole BEM, or energetic methods, such as the energy BEM/FEM, must be used. Their feasibility is to be validated using a generic car measurement setup. In addition to the numerical challenge, there are also special demands on the measurement setup in this area, as only small changes in the microphone position can lead to significant differences due to the short wavelengths.

ID:03_Dilba

Personalized sound zones technology is a groundbreaking solution for enhancing the travel experience in consumer vehicles. By creating acoustically separated areas within the car interior, this technology allows for novel use cases, such as robust individual interaction with a voice assistant for the driver and simultaneous high-quality private voice calls for passengers.

Personalized sound zones require two components: source separation at voice capture and personalized audio reproduction. This paper focuses on source separation, by analyzing how microphone types and positioning affect the starting conditions for the task and presenting a solution that improves the separation and enhancement performance for any microphone type.

ID:03_Buccoli

This paper introduces a novel approach to enhancing loudspeaker performance by employing Long Short-Term Memory (LSTM) neural networks to linearize the driving force on the voice coil. The method enables transducer engineers to design more cost-effective drivers by prioritizing the optimization of mechanical components (such as the membrane and suspension), while allowing for greater compromises and flexibility in the design of the magnetic system (including the voice coil and motor), whose deficiencies are compensated by the controller. This flexibility is particularly beneficial in challenging applications where size, cost, and weight constraints are significant factors, such as in automotive and portable devices.

We propose a control algorithm that linearizes the driving force by adjusting the voice coil current and compensating for the force factor BL nonlinearity. By focusing on the current signal, which is readily obtainable, and the BL nonlinearity, which is relatively stable and predictable, our controller architecture remains practical and robust. This intentional avoidance of complex and potentially time-varying mechanical nonlinearities and the associated costly acquisition of mechanical signals (e.g., displacement, velocity, and acceleration) ensures feasibility, especially in production environments.

The performance of the controller was rigorously evaluated using a real 2.5-inch driver deliberately engineered with compromised voice coil characteristics, resulting in a distorted BL curve and heightened levels of distortion. Comparative analysis was conducted against a non-controlled sample of a similar driver with a properly designed magnetic system (i.e., both drivers sharing the same mechanical system, but the controlled unit exhibiting suboptimal magnetic properties). Our proposed controller algorithm achieved approximately 10 dB reduction in distortion above the resonance region, approaching the performance of the well-designed unit.

ID:03_Volkov