Writing on Voice AI, speech research, and machine learning.
LLM-based ranking and reasoning algorithms for search and recommendation: why asking an LLM to rank a shortlist outperforms embedding distance on compositional queries, a research-grounded tour of pointwise, pairwise, listwise, setwise, and tournament-style ranking, and a package that implements all of them over any provider.
Taking DualTurn from an Interspeech 2026 paper into a live voice agent: a ~1.5M-parameter model trained with zero manual labels, fastest end-to-end in real-call testing against cloud and open-source turn detectors, the only one in the industry that tells a backchannel from a real interruption, and set up to keep improving for free from every call it handles.
The checklist for making a voice agent "sound human": less interrupting, more backchanneling, an even exchange, mostly targets things that aren't broken. Measured against the right human (task dialogue, not casual chat), almost everything holds up except response speed, and fixing that carelessly breaks the one thing the agent already gets right.
Speech-to-speech models know when to speak but can't reason. Cascaded LLM pipelines can reason but only react to silence. DualTurn pretrains on dual-channel human conversation to bring S2S-level turn-taking into a standard ASR-LLM-TTS stack.
A small multimodal LLM that reads paralinguistic signal — emotion, prosody, speaker traits — directly from speech audio instead of through an ASR transcript, built alongside the release of SpeechLLM at Skit.ai.