AES Europe 2026

This paper presents a systematic literature review on
inverse synthesis; sound matching, which focus on
predicting synthesizer parameters to recreate a target
audio waveform. Automating this process using machine
learning is impeded by distinct technical challenges: many
to one mappings where different parameter settings produce
the exact same sound, the non-differentiability of
commercial black box synthesizers, a scarcity of musically
structured training data,; a lack of standardized
perceptual metrics. Existing approaches are categorized
into non-differentiable synthesizer methods, utilizing
evolutionary algorithms; deep learning, incorporating
techniques to bypass gradient limitations such as neural
proxies or generative models. In contrast, differentiable
synthesizer methods, enable the integration of audio loss
functions into training pipelines via custom signal
processing environments. The analysis identifies a critical
reliance on spectral representations for evaluating
perceptual similarity, given that parameter based metrics
frequently fail to align with human hearing. The findings
indicate that while deep learning has reduced inference
times, the field lacks a unified production solution.
Future progress requires the establishment of standardized
benchmarks to evaluate models, the implementation of novel
advancements in generative models not yet applied to this
problem,; the development of hybrid architectures to
simultaneously address these distinct technical challenges.