Auralius V3 Voice Engine is now GA
Experience sub-50ms latency with our new streaming gRPC pipeline, emotionally expressive RAG capabilities, and a multi-agent swarm that works inside every call.
Today we are officially launching the Auralius V3 Voice Engine to all production tenants. After six months of architecture work and hundreds of hours of real-call testing, the new engine delivers a time-to-first-word of ~1.5 seconds — down from 4.5 seconds in V2 — without sacrificing knowledge-base accuracy or per-tenant isolation.
What changed under the hood
The biggest structural change is moving from a request/response HTTP pipeline to a bidirectional gRPC streaming seam between the FastAPI control plane and the Rust media plane. As soon as the speech-to-text engine produces a final transcript, the backend begins streaming LLM tokens downstream. The media plane accumulates tokens until a sentence boundary, then immediately feeds the first sentence to ElevenLabs TTS — while the LLM is still generating the second sentence.
RAG retrieval now runs in parallel with the opening of the LLM stream. For most queries the vector search resolves before the model has consumed even the first token of context, so the retrieval cost adds zero latency to the hot path.
Multi-agent swarm in every call
V3 ships with the first production version of per-call supervisor routing. When a tenant has more than one agent configured with a role field, the supervisor (a cheap classifier) reads the incoming utterance and routes to the right specialist — Sales, Support, Legal — before the main LLM generates a single token. The handoff is invisible to the caller and carries the full conversation context.
Background swarm signals
Alongside the main response, a lightweight background analyzer runs in parallel and emits tone and emotion signals as CallEvent rows. These appear in your Call Logs drawer as a live timeline. Emotion inference is opt-in per tenant and off by default in accordance with EU AI Act Art. 5.
How to upgrade
No migration is required. All existing agents and knowledge bases are automatically served by the V3 engine. You can verify the engine version in your call logs under engine_version: "v3". If you notice any regressions, open a support ticket and include the X-Trace-Id from the affected call.
Published August 18, 2026 · Kolerr Lab Engineering Team
