PICO-1.
A single-voice, English-only live speech-to-speech model, tuned for CPU first. Efficient generation and a short audio buffer aim to keep playback smooth through brief slowdowns.
Overview
PICO-1 entered development on August 13, 2026 as the first model in our live speech-to-speech series. It receives speech and generates spoken responses. We are working on expressiveness, responsiveness, and efficient local generation together.
PICO-1 is designed around a custom streaming system. As PICO generates audio, it produces it in short pieces called frames. Those frames pass into a buffer that holds them in order until the previous frame has played. Playback uses a short reserve of generated audio, and the buffer keeps handing frames to the output in sequence as PICO keeps producing.
The buffer primarily smooths brief generation or compute hiccups. Playback draws on the reserve during a slowdown; efficient generation can catch up and refill it afterward, making the interruption less audible. A sustained slowdown can still exhaust the reserve, so this is not a guarantee of uninterrupted audio under every workload.
Since we set a clear order of priority across projects, PICO-1 continues at a reduced pace, last in line behind Kova, Deriva-1, and Hikaru-1. See Evaluating scope for the full reasoning.
Planned PICO-1 capabilities.
Time-to-first-audio figures are current targets, not guarantees. Actual results vary with hardware, input length and complexity, and the rest of the workload on the machine.
Not published yet.
Benchmark runs are not finished. We would rather leave this table empty than publish numbers we can't stand behind, so every score below is marked N/A until that changes.
| Benchmark | PICO-1 | Comparable speech-to-speech models | Comparison |
|---|---|---|---|
| Naturalness (MOS) | N/A | N/A | N/A |
| Measured time-to-first-audio | N/A | N/A | N/A |
| Word error rate (intelligibility) | N/A | N/A | N/A |
| Stability under load | N/A | N/A | N/A |