45ms STT / 38ms TTS
The low-latency speech engine for real-time voice agents.
Streaming speech-to-text and text-to-speech with sub-50ms round-trip, so voice agents answer at human conversational tempo instead of leaving callers waiting. Bootstrapped, built by engineers, now in early access.
Based on internal benchmark tests across 1,000+ audio samples in English and Spanish. Results vary by network and audio quality.
530ms gaps kill real conversations.
The average voice AI pipeline today takes 530ms from end of speech to start of response. That pause is long enough for a caller to assume the agent froze, or to start talking again, or to hang up entirely.
The delay compounds across three layers: a speech recognition model that buffers audio before transcribing, a language model that takes time to generate tokens, and a text-to-speech model that renders the full sentence before streaming. Each layer adds latency that the next layer inherits.
Reducing latency in any single layer helps but does not solve the problem. Smallest AI compresses all three hand-offs simultaneously through streaming-first architecture, co-located inference, and quantized acoustic models tuned for conversational phoneme patterns.
A 530ms silence in a phone call reads as a dropped call or a confused agent. At 133ms, the gap falls inside normal human response time and the conversation feels continuous.
Three layers, one streaming pipeline.
Each stage begins processing before the previous one finishes. Audio frames feed the STT model as they arrive. Tokens feed TTS synthesis as the language model produces them.
Streaming STT
45ms median
Audio frames are transcribed as they arrive. The model emits partial transcripts immediately rather than waiting for an end-of-utterance boundary. Final transcript confidence threshold is configurable.
Token passthrough
50ms handoff
Text tokens from the STT model are forwarded to the language model without buffering. The language model begins generating a response before the transcript is complete, using a lookahead confidence window.
Streaming TTS
38ms median
TTS synthesis starts on the first semantic phrase boundary. Audio bytes are streamed to the caller as they are rendered, not after the full sentence is complete. Natural prosody is preserved across chunk boundaries.
See the difference in a 10-second voice turn.
Same language model, same network. The difference is the speech layer architecture.
Pipeline segments proportional to measured latency. Proc = LLM first-token time on a 15-token prompt, 4o-mini equivalent.
Built for teams that ship voice agents, not experiment with them.
Contact center AI teams
Replacing IVR trees with conversational agents that handle interruptions, accents, and background noise. Latency at human tempo reduces caller escalations caused by perceived unresponsiveness.
Voice AI infrastructure builders
Embedding speech into a larger platform and need the speech layer to stay out of the way. Drop-in WebSocket API with OpenAI-compatible response format. No model fine-tuning required to start.
AI-first product companies
Launching voice-first features where the speech layer needs to match the product quality bar. Customizable voice identities, real-time word error rate telemetry, and language-specific acoustic models.
Applied AI labs
Running latency-sensitive voice experiments and benchmarks where the speech layer should not be the variable. Access to raw frame-level latency logs and per-segment trace export.
We dropped caller abandonment by 22% after switching. The difference is audible. Callers stop second-guessing whether the agent is still there.
Engineering lead, a contact-center automation team in financial services
Early access program, June 2026. Shared with consent, name withheld.