Sub-80ms Voice Pipeline

OP26X is designed around a continuously streaming media path. Audio is processed as it arrives instead of waiting for a complete turn, while dialogue state and knowledge retrieval remain available to the agent.

PSTN → carrier media → VAD and STT → dialogue state → fast-path reasoning → streaming TTS → media gateway → PSTN

Pipeline stages

1. 01

   Carrier media ingress

   The telephony gateway receives streaming audio and call-control events from the active carrier path.

2. 02

   Voice activity detection

   Speech boundaries let the agent distinguish a completed turn from a pause and yield immediately when the caller interrupts.

3. 03

   Streaming speech recognition

   Speech-to-text results update the active dialogue turn as confidence improves.

4. 04

   Dialogue orchestration

   Conversation state, agent instructions, call variables, and approved knowledge are assembled for the next response.

5. 05

   Reasoning and action

   The model can answer, ask a clarification, invoke a configured action, or transfer to a human. Fast paths can avoid unnecessary model work for common cases.

6. 06

   Streaming speech synthesis

   Text is synthesized and delivered in chunks so playback can begin before the full response is complete.

7. 07

   Post-call processing

   The platform finalizes the transcript, disposition, qualification result, recordings, and subscribed events.

Latency-sensitive design choices

• Keep the first message short so playback can start early
• Ask one focused question per conversational turn
• Preload approved property and project context before dialing where possible
• Use deterministic fast paths for greetings, confirmations, and common transfers
• Avoid large real-time knowledge documents and unrestricted external lookups in the critical path
• Measure turn latency by network, language, model, carrier, and prompt rather than by one global number

Interruption handling

When a configured pause-on-speech policy detects caller speech, playback should stop before the current response completes. The caller utterance then becomes the new active turn, and partial output from the interrupted response should not be committed to conversation state.

Operational signals