Duet

Most production voice agents still run a cascaded pipeline — speech-to-text, LLM, text-to-speech — glued together with "wait for silence, then respond." It works, and it feels like talking to a walkie-talkie. Duet gives an existing cascaded stack genuine full-duplex behavior: backchannels ("mm-hm"), mid-sentence interruption recovery, responses in under ~300 ms — without the platform rebuilding anything.

The part I care most about is the benchmark. The full-duplex core (Moshi) and a cascaded baseline ran the same ten SDR scenarios against the same simulated caller. Duet's median response handoff came in at 240 ms against the cascade's 1,880 ms — roughly 8×. It also backchannels while you talk, which the cascade cannot do at all.

And then the numbers that hurt: Duet's takeover rate is worse (0.24 vs 0.00) and its p95 tail is slower. Both went in the published table. A benchmark that can't embarrass you isn't a benchmark; it's marketing. Reasoning quality is CI-gated at a 90% golden-eval floor (currently 97.6%), and the observability stack — Langfuse, Prometheus, Grafana — ships in the repo.