CVSSP & PAI External Seminar

Abstract
Audio understanding aims to enable machines to recognize and interpret acoustic events and scenes in everyday environments. An important direction in this field is automated audio captioning, which generates natural language descriptions of audio content and connects audio understanding with natural language processing. Recent advances in deep learning and large-scale audio-text datasets have led to significant progress in this field, supporting a growing range of applications and downstream tasks. In this talk, we will introduce our research in audio understanding, starting with early approaches to audio captioning, including transfer learning, diverse caption generation, and the Audio Captioning Transformer (ACT). We will then discuss WavCaps, a large-scale audio captioning dataset, and its role in audio-language learning, followed by our more recent work on SLAP. Finally, we will discuss open challenges and possible future research directions.

Speaker Bio
Xinhao Mei is a Research Scientist at Spotify, where he works on audio and music understanding and generation. He received his PhD from the University of Surrey in 2024 under the supervision of Prof. Wenwu Wang. Before joining Spotify, he was a Research Scientist at Meta Reality Labs. His research interests include audio understanding, audio-language modelling, and generative models for audio and music.

Abstract
How can machines listen reliably when voices overlap, sound sources move, and acoustic environments change? Drawing on Dr Hanyu Meng’s doctoral research at the University of New South Wales, this talk explores biologically inspired adaptive audio front ends for robust machine hearing. It presents approaches to adaptive audio representation, binaural selective listening for speaker localisation and tracking, and acoustic context estimation from spatial audio, highlighting the key ideas, challenges, and lessons learned. The talk also provides a brief overview of her ongoing research in music transcription and audio-visual room impulse response (RIR) prediction, before concluding with open research questions and opportunities for collaboration across speech, music, and spatial audio.

Speaker bio
Hanyu Meng received her PhD in Electrical Engineering from the University of New South Wales (UNSW Sydney) in 2026. She is currently a Research Scientist Intern at Meta Reality Labs in Cambridge, UK. Her research focuses on biologically inspired audio front ends, binaural selective hearing, speech enhancement, spatial audio, and music information retrieval. She has also interned at Dolby Laboratories, Sony Corporation, and Sony Computer Science Laboratories, working on acoustic context estimation from spatial audio, instrument sound synthesis, and audio editing.

She received the 2026 Australasian Speech Science and Technology Association (ASSTA) New Researcher Award. She actively contributes to the speech, audio, and music information retrieval communities, serving as a reviewer for ICASSP, Interspeech, ISMIR, and journals including IEEE/ACM Transactions on Audio, Speech, and Language Processing.