Decomposing audio recordings with artificial intelligence

In the music world, artificial intelligence (AI) can now generate new sound clips, recommend music to listeners, and aid academic analysis of music recordings. Central to this advancement is a novel use of one-dimensional (1D) filter convolution, a decades-old technique broadly used in fields like sound engineering and telecommunications. Traditionally, 1D convolutional filters—small patches of signal that slide over larger signals to extract features—were manually predefined for specific purposes, such as muting frequencies above a threshold. But in audio-based AI systems, they are learned by the AI, making them both more opaque and versatile than their handcrafted counterparts. As AI music systems become more popular, the use of 1D filter convolution has expanded. Yet, the technique remains understudied in music scholarship. This talk shows how musical AI systems leverage 1D convolutional filters to decompose audio into short-term spectral features, suggesting potential new tools for sound analysis. While I will use musical examples, the implications extend beyond music. For example, AI-learned filters could enhance our ability to understand and manipulate environmental sound or improve sound processing in devices like hearing aids and smartphones. Ultimately, this talk aims to challenge existing methodologies and encourage innovative research in acoustics and related fields.