Plan a migration from cascading voice to native audio

Inventory behavior rather than copying model names

A cascading voice path has explicit recognition, conversation and synthesis stages. Native audio sends and receives audio through the conversation provider, which can manage those stages internally. Before migration, list the behaviors your application depends on: turn timing, language switching, interruption recovery, tool approvals, transcript retention and failure handling. A shorter component list is not evidence that those behaviors are identical. Keep the same owned number and workspace task boundaries during evaluation so differences can be attributed to the media change rather than simultaneous changes in identity and permissions.

Check the captured media mode and transport

Lupin’s shared stack contract distinguishes the streaming cascade from native_audio even when the stored transport remains scribe_ws for compatibility. The native path names Eleven Realtime v1 Max while recognition and synthesis can be managed components. The transport label alone is therefore an insufficient migration flag. Read the captured media mode and actual diagnostics for each evaluated call. Do not claim that a native session still uses a standalone Scribe recognizer or V4 synthesis socket merely because those components appear in the cascade’s configuration.

Preserve workspace tool authority

The conversational interface can change while Lupin’s workspace agent continues handling private tools through its existing backend and approval rules. Keep authenticated callback requirements, scoped credentials and exact pending approvals intact. Spoken assent during a call does not execute a pending sensitive action. A migration acceptance test should include a request to read private workspace content and another to perform a consequential write. Verify that the new media path reaches the same authorization boundary instead of accidentally treating conversational confidence or caller ID as an access credential.

Translate metrics and fixtures deliberately

A cascade exposes recognition commits, text completion and synthesis intervals. Native audio can expose different provider events and managed timing scopes. Define the interval you will compare before collecting results, and leave fields unknown when the new path does not measure them. Reuse synthetic dialogues for pauses, interruptions, mixed-language names and a tool result followed by correction. Include late audio chunks and a provider disconnect. A migration is successful when the intended interaction remains correct under these cases, not only when the fastest reply sounds more natural.

Roll out with a reversible routing decision

Keep admission limits, allowed peers, the monthly seconds allowance and maximum duration fixed while evaluating the new path. Have the operator validate native audio separately from SIP or cascade setup, then capture the chosen stack on each admitted call. Retain a tested route back to the previous media mode and avoid changing number ownership during the same rollout. Record unresolved gaps such as missing measurements or transcript differences. This plan establishes migration readiness without dialing a real number or authorizing a live provider change as part of the written guide.

Define the rollback threshold

Before rollout, decide which regressions require returning to the prior path: lost correction context, stale audio after interruption, failed transcript binding or missing approval prompts are concrete examples. A modest latency improvement should not override those failures. Keeping this threshold explicit prevents a more natural sounding sample from masking an authority or lifecycle regression.

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.