Full duplex audio does not identify the conversation model

Name the transport behavior precisely

Full duplex describes the ability to carry incoming and outgoing audio at the same time. It lets a system keep listening while the assistant is speaking, which is necessary for natural interruptions. It does not name a speech recognizer, synthesis model or reasoning backend. A product can have duplex transport and still feel slow because it waits too long to finalize a turn. Conversely, a fast text model cannot make a one-direction-at-a-time audio path support interruption. Describe these properties separately when documenting or evaluating an agent.

Follow the actual processing graph

In a cascading design, audio becomes text, a conversation component chooses a response, and synthesis turns the response into outgoing audio. A native audio design accepts and returns audio through the conversational provider while retaining separate workspace tool boundaries. Lupin captures the configured stack at call admission so operators can inspect the actual path later. The current gateway cascade names Scribe v2 Realtime, Eleven Realtime v1 Max and V4 Turbo. Native audio has managed provider components; the streaming transport name alone cannot identify which media mode ran.

Keep workspace intelligence distinct from spoken delivery

When a conversation needs workspace operations, Lupin delegates through its existing agent backend and permission system. That backend remains responsible for tool discovery, approved actions and private workspace context. The phone voice is not evidence that the workspace agent changed to another named language model. Read the backend diagnostic fields for observed model and provider information when available, and leave unavailable measurements unknown. This distinction makes a useful incident report possible: the conversational response may be quick while an approved external application request is the slower part.

Evaluate duplex behavior with an interruption case

Prepare a synthetic dialogue where the assistant starts a long answer and the caller begins a short correction halfway through. Observe whether input audio continues to arrive, whether queued output stops, and whether a stale response resumes after the correction. Successful transport overlap alone is insufficient: the conversation context must also recognize that the earlier response was interrupted. Repeat with background noise and a brief acknowledgement so the system does not mistake every sound for a new instruction. Record the event ordering alongside any resulting transcript.

Use separate claims in a capability description

A precise capability statement names the enabled transport, media mode, observed interruption behavior and captured model configuration. It also states the scope of testing: a local fixture cannot prove handset audio quality on a real carrier path. Avoid phrases that imply a provider controls every part of the agent or that duplex audio guarantees human timing. When comparing two implementations, hold the dialogue, voice and network conditions steady and change one component at a time. That produces evidence about the actual bottleneck instead of attributing every difference to a model name.

Keep microphone evidence narrow

A local browser microphone test can confirm that capture continues during generated playback. It cannot establish that a carrier path, SIP deployment or private workspace tool bridge works. Label that result as capture overlap, then use separate fixtures for cancellation and authorized tool routing before describing the complete phone experience.

Sources & verification

Reviewed against current checked-in Lupin routes, shared contracts and official provider references. No live calls, OAuth authorizations, app actions or purchases were performed.

• Lupin phone, SMS and calls ↗Checked 2026-11-05
• ElevenLabs model reference ↗Checked 2026-11-05

Prepared with AI assistance by Lupin editorial. How we research and maintain these guides.