Skip to main content

Voice capacity

How many conversations can run at once depends on the GPU that runs speech recognition and the text-to-speech voice. The number is measured, not configured: STELLA has no software cap on sessions, and the GPU is what slows the voice down first.

Where to see it​

Settings → Admin → Voice capacity shows, for each server and GPU, the number of simultaneous conversations that stayed fluent, the GPU it was measured on and the date, with a table of what each tested level looked like. "At least N" means the test stopped at its highest level without finding the limit.

Until a load test has been run on a server, the card says "Not measured yet".

Measuring your server​

Run the load test in scripts/load-test/. It simulates more and more simultaneous sessions until speech recognition or the voice degrades, and can publish the result to the dashboard. See scripts/load-test/README.md for the setup, what counts as "degraded", and the options.

Measure on the GPU you deploy on and without other load. A GPU that is faster or slower than the one measured will carry a different number.

Known limitations​

  • Development server (Tesla T4), Qwen3 voice, streaming playback: measured capacity is 0. One session already starves the voice (9.7% of playback against the 5% limit, judged with an 800 ms player pre-roll), and two sessions collapse: speech-recognition finals take about 7 s and the voice's first audio about 6 s. The GPU is at about 82% with one session and about 97% from three. The shared Qwen3 model is the ceiling.
  • This measurement is of STELLA_TTS_PLAYBACK=stream specifically, forced for the whole session. Since #669, the default (auto, and unset) no longer stays in stream on hardware this slow: it switches to sentence-by-sentence playback after a few sentences measure slower than real time, which rules out Playout starved for the rest of the session but not the underlying GPU contention above. Pending a live measurement on the development server after deploy — not yet re-run in auto.
  • This says nothing about other GPUs (production's has not been measured) or about the sentence-by-sentence playback mode (STELLA_TTS_PLAYBACK=sentence), which the test does not model. Capacity also does not include the language model, LiveKit or agent CPU.