PICO-1 enters development.
Work begins on PICO-1, our live speech-to-speech model. A single voice, English-only, tuned for CPU first, with efficient generation and a short playback buffer to smooth brief compute hiccups.
Clarification, September 27, 2026: PICO-1 is a live speech-to-speech model. This article previously described it incorrectly as text-to-speech. We have corrected its scope and explained the buffer more precisely. Development now continues at a reduced pace; see the current model page.
At Sential Innovations, our work has always been about shipping the best models and integrations we can. To that end, we have started development on a new series: PICO. The first model in the series is PICO-1, our custom live speech-to-speech model.
PICO receives speech and generates spoken responses. We are working on expressiveness, responsiveness, and efficient local generation together. Time to first audio is one of our development targets.
The streaming buffer
PICO is designed around a custom streaming system. As PICO generates audio, it produces it in short pieces we call frames. Those frames pass into a buffer that holds them in order until the previous frame has been played. Playback uses a short reserve of generated audio. PICO keeps producing; the buffer keeps handing frames to the audio output in sequence.
The buffer mainly softens brief generation or compute hiccups. Playback can draw on audio already generated while the model slows down, and efficient generation can catch up and refill the reserve afterward. The goal is to make short hiccups less audible. Sustained slowdowns can still empty the buffer and interrupt playback.
Scope for the first release
PICO-1 is a single-voice model. The planned first release targets English only. Language support in other models integrated with Kova does not imply language support in PICO.
We plan to make PICO-1 available locally through Kova. Our current targets for time-to-first-audio (TTFA) are:
- GPU: 50–600 ms
- CPU: 300 ms–1.4 s
These are ranges, not guarantees. Actual rates will vary with the hardware you have available, the length and complexity of the input, and the rest of the workload on the machine. PICO is tuned for CPU first and GPU second, the opposite of how most of our other models are tuned, because the kind of device that benefits most from a local voice model is often the kind of device that does not have a discrete GPU.
What comes next
Development is underway. The near-term plan is to finish the PICO-1 training and integration work, validate TTFA on a range of hardware, and ship the model through Kova once we are confident in both the output quality and the streaming behaviour. We will share more when there is something concrete to report, not before.