Sound matching consists in generating a sound that resembles or replicates a given target sound.

This problem requires adjusting the parameters of a chosen synthesis model to achieve this result. It is typically addressed by optimizing objective functions designed to evaluate the properties of the estimated solutions.

Modern optimization methods often rely on the gradient, which represents the vector of partial derivatives of the objective function with respect to its input variables. By iteratively adjusting the model parameters in the direction indicated by the gradient, these methods converge toward a solution. While highly efficient in general machine learning, gradient-based methods are less suitable for audio optimization tasks. Many such problems often involve non-differentiable models and objective functions, gradients that are challenging to evaluate and complex optimization landscapes with numerous local minima.

To address these issues, this internship will focus on exploring gradient-free optimization methods, a class of algorithms that do not rely on the derivatives of the objective function. Many of these methods are biologically inspired, such as the Genetic Algorithm, which mimics evolutionary processes like mutation and selection, and Particle Swarm Optimization, which models the collective behavior of organisms like bird flocks and fish schools to generate solution estimates. As some of these techniques have shown to be promising in the context of sound matching with synthesizers or sound effects, their further exploration could lead to more efficient, robust, and effective approaches to solving this problem.

• Conduct a bibliography on gradient-free optimization methods to establish a thorough understanding.

• Implement the considered methods and set up a protocol to evaluate their performance, efficiency and robustness.

• Assess the quality of the resulting sounds.

• Compare the most effective methods to the gradient-based approaches integrated into our internal tool for parameter optimization.

• Master-level or engineering school student (final year).

• Strong interest in research.

• Advanced knowledge in Digital Signal Processing (DSP).

• Proficiency in Python and previous experience with scientific libraries.

• Good command of English.

• Experience with C++ programming, machine learning frameworks (such as PyTorch, Keras, or TensorFlow), audio DSP, good mathematical reasoning and communication skills are a strong plus.