Ambisonics is a versatile framework for spatial audio capture, encoding, and reproduction, with growing adoption in industry as the standard format for immersive spatial audio. Ambisonics is based on the spherical harmonics representation of sound fields, with its signal processing having evolved from classical spherical array theory to modern machine learning. This talk reviews Ambisonics encoding approaches as a means to generate Ambisonics signals from microphone recordings, first for spherical arrays, then for wearable arrays and arrays of unconstrained geometry using methods ranging from linear Ambisonics Signal Matching to neural network-based approaches that implicitly leverage information in the audio data. Neural network-based approaches, once incorporated into spatial audio applications, require loss functions that reflect auditory spatial perception, a topic of growing interest that also opens new challenges. These perceptual loss functions are particularly relevant to recent work on Ambisonics upscaling - moving beyond first-order to higher-order representations. Beyond upscaling, Ambisonics further extends to speech enhancement, both for improving spatial audio quality and, more broadly, for enabling array-agnostic processing through its modular, geometry-independent structure. Together, these topics illustrate the growing role of Ambisonics as a unifying representation in spatial audio and point to promising directions for future research.

The ability to navigate complex acoustic environments depends on more than recognizing sounds. It requires identifying what is behaviorally relevant, organizing sensory inputs into meaningful objects, and adapting to changing goals and contexts. This talk will examine the computational principles that support these abilities, drawing on studies of auditory attention, predictive processing, and auditory scene analysis. I will present behavioral and neural evidence illustrating how top-down goals, statistical learning, and bottom-up organization interact to shape auditory representations and guide selective listening in natural acoustic scenes. I will conclude by discussing how these same principles are beginning to influence modern audio systems, offering opportunities for closer integration between auditory neuroscience and machine listening.

Large audio-language models (LALMs) are emerging as a powerful approach for addressing a broad range of audio signal processing problems, including generation, captioning, editing, and source separation. By integrating large language models (LLMs) with audio modeling techniques, they enable more flexible and intelligent audio systems. This talk focuses on language-queried audio source separation (LASS), a recently introduced framework that extracts desired sounds from an audio mixture using natural language queries. LASS offers an intuitive and scalable interface for applications such as automated editing, remixing, and audio rendering. We start by outlining the problem and its motivation, making a connection with conventional paradigms for speech and universal source separation. We then present two LASS methods: AudioSep and FlowSep. AudioSep is a text-driven model that uses query and separation networks to estimate time–frequency masks and extract target sounds. FlowSep is a generative approach based on rectified flow matching (RFM), modeling transitions from noise to target features in a variational autoencoder (VAE) latent space and reconstructing audio via a decoder and vocoder. We also present datasets, evaluation metrics, experimental results, and sound demos. Finally, we link source separation to audio editing and generation, highlighting recent advances (e.g. the RFM-Editing, AudioLDM, and AudioLDM2 models), and conclude the talk with future research directions in this evolving field.