Zhizheng Wu

FlexiSLM can operate at frame rates as low as 4.0 Hz while maintaining high-quality speech, effectively halving inference time compared to traditional models.

PupuJEPA reveals that leveraging 2D spectrograms can dramatically enhance music representation learning, outperforming traditional 1D models across multiple tasks.

Speaker Anonymization enables real-time voice conversion without the latency penalties of future context buffering, revolutionizing streaming applications.

Paralinguistic cues can be effectively harnessed in dialogue systems, leading to a 175% improvement in safety response accuracy without compromising overall model performance.

Geometry, not token discreteness, is the key to unlocking superior performance in speech-to-LLM integration.

Achieve state-of-the-art few-shot RIR prediction by explicitly modeling the connection between geometric features and the RIR power spectrum.

SLMs that seem safe with text inputs can completely fail when the same content is spoken, revealing a critical "speech grounding gap" in current models.

Forget complex disentanglement architectures or low-quality synthetic targets: MimicLM achieves superior voice imitation by cleverly using synthetic speech as the *source* and real speech as the *target* in a pseudo-parallel training setup.