LEMAS LEMAS: A 150K-Hour Large-scale Extensible Multilingual Audio Suite with Generative Speech Models
arXiv:2601.04233, 2026
I am a Senior Research Engineer at XPENG Robotics, responsible for speech interaction on the IRON humanoid robot. My research interests include generative speech/audio, speech-native interaction for embodied robots, and natural-language interfaces for controllable music creation.
Previously, I worked at IDEA Research, UBTECH Robotics, and SF Tech, and interned at Samsung Research China - Beijing. I received my M.Sc. in Robotics from King's College London and my B.Sc. in Electronic Information Science and Technology from Central South University.
Publications
LEMAS 
MelTok 
STFTCodec STFTCodec: High-Fidelity Audio Compression through Time-Frequency Domain Representation
IEEE ICME, 2025
VC Open Source Projects
Speech Dataset + Synthesis Models
LEMAS
logo 150K+ hours, 10 languages, word-level timestamps, plus TTS/Editing Models.
Speech/Audio Demopage
VoicePlatform
logo Audio/video AI workflow for ASR, TTS/VC, speech editing, dubbing, and translation.
AI Music LiveCoding Platform
SVLive
logo Natural-language live coding with Strudel, Vital, and editable synthesis.
Video Dataset
OpenVideo
logo 672+ hours / 106K+ clips / 720p text-to-video dataset with annotation and validation tooling.
Explorations
Most interactive systems today are still largely passive instruction followers.Next-generation HMI systems require a shift from discrete commandsto continuous understanding across time, modalities, and turns.Perception -> Understanding -> Executionperception is not detection, but prediction;understanding is not classification, but state tracking;execution is not synthesis, but externalized strategy.
From Perception to Action
A Closed-Loop Architecture for Next-Generation Human-Machine Interaction Systems
let cpm = 28.0;await samples('.../dreamy-demo-vital/strudel.json')let KEY = 'vital_keystation';let PAD = 'vital_flute_1';let HORN = 'vital_damped_horn';let track_melodies = arrange([10, track_melody1][16, track_melody2][16, track_melody3]);stack(track_arp_accompaniment, track_melodies).cpm(cpm);
SVLive
Dreamy Strudel live-coding demo with Vital sample presets. Opens directly in strudel.cc.
To Be Continued