Han Han 
about      ︎︎︎︎


Research 

mersenne

This work constructs a monochord string bench and records a dataset of pluck string sounds of different cultures, materials, lengths, tensions etc. Simultaneously a machine listening model is trained on synthetic data from a string physical model, to extract physical parameters. Now, how would the machine listening model perform on the real world sounds?
This work addresses the problem of sim2real and domain adaptation in face of shift in query data’s distribution.
ongoing

towards constructing a
historically-grounded gesture-timbre space of Guqin playing techniques



This work curates a set of non-composite playing techniques from the 9th century treatise “Tangchenzhuo zhifa”, also extracting from which key gestural descriptions of each technique. A dataset of ~4 hr audiovisual materials of each isolated techniques is recorded at the Laboratoire Acoustique Musicale (LAM), Sorbonne University in aug. 2023. This work then examines the influence of various gestural degrees of freedom on guqin timbre, by means of isomap, listening tests and others.

invited artists:
Li-yu You 游丽玉: guqin master and scholar
Cixian Lu 陆慈娴: composer

presented as “work in progress” at Timbre  workshop in Thessaloniki, Greece, aug. 2023

poster ︎


presented at Journée de restituion under the project call “collaborations entre jeunes chercheuses et artistes” supported by l’association française d’informatique musicale (AFIM) and la société française d’ethnomusicologie (SFE) in bibliothèque La Grange Fleuret, Paris, dec. 2023
slides ︎


perceptual-neural-physical sound matching

Sound matching is the task of finding the optimal set of parameters to a sound synthesizer such that a target sound is matched as closely as possible. Two common learning objectives in deep learning approaches are spectral loss and parameter loss. The former being perceptually-relevant yet the resulting optimization landscape has lots of local minimas. The latter is fast and easy to reach convergence yet is not perceptually-driven. A new loss function (PNP loss) is proposed in this work. PNP loss is the optimal quadratic approximation of spectral loss while being as fast as parameter loss. 
presented at c4dm seminar in Queen Mary College of London, nov. 2022 slides ︎
paper  ︎

presented at ICASSP 2023, Rhodes Island, jun. 2023



wav2shape  


wav2shape aims to extract physical parameters from waveforms of percussive sounds synthesized by a rectangular drum physical model. 
presented at Forum Acousticum 2020 
paper︎
code ︎

Publications


Loïc Jankowiak, Han Han, Vincent Lostanlen, Mathieu Lagrange, “Towards multisensory control of physical modeling synthesis”. Internoise 2024.

Han Han, Vincent Lostanlen and Mathieu Lagrange, "Learning to Solve Inverse Problems for Perceptual Sound Matching," in IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 32, pp. 2605-2615, 2024

Vincent Lostanlen, Daniel Haider, Han Han, Mathieu Lagrange, Peter Balazs, Martin Ehler, “Fitting Auditory Filterbanks with Multiresolution Neural Networks”. IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA), 2023
 
Han Han, Vincent Lostanlen, Mathieu Lagrange, “Perceptual Neural Physical Sound Matching”. International Conference on Acoustics, Speech, and Signal Processing (ICASSP), 2023
 
Cyrus Vahidi, Han Han, Changhong Wang, Mathieu Lagrange, György Fazekas, Vincent Lostanlen, “Mesostructures: Beyond Spectrogram Loss in Differentiable Time-Frequency Analysis”. Journal of the Audio Engineering Society 2023. 

John Muradeli, Cyrus Vahidi, Changhong Wang, Han Han, Vincent Lostanlen, Mathieu Lagrange, George Fazekas. “Differentiable time-frequency scattering in kymatio”. International Conference on Digital Audio Effects (DAFx), 2022. Best paper award.

Han Han and Vincent Lostanlen, “wav2shape: Hearing the shape of a drum machine”. Forum Acousticum 2020. Best student paper award.