New musical instruments change how music is made and experienced. Electric guitars enabled rock, and synthesizers enabled techno. I focus on what the next instrument could be in the AI era: interactive music AI systems that people can actually play.
My research centers on autoregressive generative models and neural audio codecs, with an emphasis on real-time, streamable architectures. I work on enabling low-latency generation under practical constraints, optimizing models for interactive use and real-time deployment rather than offline benchmarks alone.
Recent work includes TokenSynth, a codec-based autoregressive neural synthesizer that enables intuitive timbre control through audio or text prompts, supporting zero-shot instrument cloning and text-to-instrument synthesis. Another project, Sori, is a real-time system that transforms arbitrary, even non-musical audio streams into symbolic music, with an explicit and controllable trade-off between input accordance and musicality.