Research
Measurements from inside a phone call with a voice agent.
Latency, turn-taking, recognition, synthesis and testing, measured on real calls and at load, explained step by step.
Voicecore: Rust media backend, road to infinite scalability with no idle GPUs
The Python media path used up the processor of one bot process at about 96 conversations. Voicecore, written in Rust, carries 224 with the GPU out of the way.ReadOne Server Unit: 100 simultaneous calls, load-tested to 176% of rated capacity
We loaded one Server Unit from a single call to 176% of its rated 100 calls. Replies held the Standard grade up to rated capacity, and every call connected.ReadOasis Speech 2: recognition in 48 languages, before and after
Oasis Speech 2 lowers the error rate on short replies by 44% across the 48 languages it recognizes, and all 48 improve.ReadListening while talking, and what the agent does with the caller's open channel
How a Dialogist agent keeps listening while it speaks, starts its reply inside the pause and tells an interruption from a cough.ReadThe shortest answers are the hardest to recognize, and the most common on a call
Most of what a caller says is a few words, and a few words are the hardest case for speech recognition. Oasis Speech 2 cuts errors on one-word replies by 49%.ReadHow we grade a conversation between a voice agent and a simulated caller
Polygon, our test bench, grades a voice agent on whole conversations, with simulated callers, eight weighted criteria, hard gates and behavior probes.ReadCadence: the workbench where we fine-tune, evaluate and gate the models behind Oasis Speech
Cadence is where we fine-tune and evaluate the models behind Oasis Speech. People and coding agents use one API, and every model must pass a gate.ReadWhat happens to a sentence after the voice model speaks it
Between the voice model and the caller, the agent trims, resamples, checks and caches every sentence, and re-checks the cache at night.ReadKnowing when the caller has finished
Across nearly 100,000 caller turns, our end-of-turn model closed most turns before the silence timer and gave back about 0.3 s per turn on real calls.Read
Next step
Get early access to Dialogist voice agents: we invite teams from the waitlist as capacity opens.
Join the waitlist
We're letting teams in gradually. We'll email you when your spot opens.