Voice AI, measured in the open.
The state of speech-to-speech, full-duplex, and audio foundation models — latency, releases, corpora, and the benchmarks that keep score. Curated by hand, every claim linked to a primary source.
First response, in milliseconds
Vendor-disclosed or community-measured time to first audible reply. Lower is better.
Model releases per half-year
When the 36 tracked models and platforms first shipped.
Recently updated benchmarks
The scoreboards that moved last. Full landscape on the benchmarks desk.
| benchmark | tier | setting | updated |
|---|---|---|---|
| SpeechJBB | native | lab | 2026-06 |
| LALM Jailbreak Taxonomy & Eval | native | lab | 2026-05 |
| AIA — Acoustic Interference Attack | native | lab | 2026-05 |
| Full-Duplex-Bench v3 | native | live | 2026-04 |
| Artificial Analysis Speech Arena | adjacent | arena | 2026-04 |
| Big Bench Audio | native | lab | 2026-03 |
Largest catalogued corpora, in hours
Log scale — the pretraining giants and the interaction sets live orders of magnitude apart.
Essays and field notes for people too busy to read every paper.
How the field measures voice, grouped by tier and setting.
Speech-LMs, realtime APIs, agent platforms, and frameworks.
The corpora behind the models, from Switchboard to the frontier.
Latest dispatches
all posts →Live on Discord, weekly in your inbox.
Fullduplex is, as the name implies, a two-way channel. Lab news and show & tell in real time on Discord; the weekly report by email; new entries and corrections through the community board.