Perceptual representations, controllable music generation, and multimodal audio — with papers, code, and listening demos.
Contributed to Qwen3.8-Omni-Flash training, focusing on agentic audio understanding and REAPER-based tasks for automatic arrangement, mixing, and podcast editing.
My contribution. Contributor during my Qwen Omni internship; worked on model training for agentic audio understanding and REAPER-based editing tasks, including automatic arrangement, mixing, and podcast editing.
A complex-spectral autoencoder that directly models frequency, phase, and stereo structure to address lost high frequencies and collapsed stereo images.
My contribution. First-author work on frequency-aware autoencoding and perceptually guided stereo refinement.
Research outcome. On the 546-track Song Describer Dataset, the model reports the best point estimates on five of seven reconstruction metrics in the paper’s comparison.
An end-to-end video-dubbing model that generates multi-speaker dialogue and coherent sound effects from uncropped video.
My contribution. Co-first author of the work on end-to-end multi-speaker video dubbing.
Research outcome. Supports multi-speaker dialogue dubbing and joint speech–audio generation, with listening demos and two evaluation benchmarks.
Music generation from text and lyrics, plus cover-song generation. My work focuses on the audio decoder, listening evaluation, and quality-oriented data curation.
My contribution. Core contributor to decoder development and audio-quality evaluation; second-listed author of the technical report.
Research outcome. The technical report presents text-to-music and cover-song generation with generative stereo rendering.
Separate vocal and accompaniment token streams preserve source structure for controllable music generation.
My contribution. Co-author; contributed the core research idea for source-aware music tokenization.
Research outcome. Enables fine-grained structural control and source-separated generation in large-scale autoregressive music systems.
Reconstructing music with perceptual weighting, phase supervision, and stereo-aware losses to preserve high-frequency detail and spatial cues.
My contribution. First-author work on perceptually grounded objectives for music reconstruction.
Research outcome. Evaluated at 44.1 kHz against open-source audio autoencoders, with improvements in spectral reconstruction and stereo characteristics.