Plan a migration from cascading voice to native audio
Preserve phone identity and approvals while comparing media contracts, diagnostics and rollback for a native audio transition.
Read captured synthesis configuration and compare provider model trade-offs without replacing the workspace agent’s identity.
A voice identifies a configured sound, while a synthesis model controls how text becomes audio. They are related but distinct choices. A voice available in a provider library is not evidence that every model supports it equally, and a familiar voice does not establish the underlying conversation model. Keep voice configuration private where it contains customer-specific identifiers. Document language suitability through provider metadata and actual evaluation rather than guessing a speaker’s gender, personality or geographic background from a name or phone number.
The checked streaming cascade uses eleven_v4_turbo synthesis, with Scribe v2 Realtime recognition and Eleven Realtime v1 Max conversation. Other supported transport or media modes can expose provider-managed components instead. Flash and Multilingual are therefore useful provider model comparisons, not universal descriptions of Lupin’s live call stack. Open Stack & timings for the call’s captured configuration and distinguish configured components from models observed in provider diagnostics. If a field is unavailable, retain that uncertainty rather than filling it with the model you expected to find.
ElevenLabs describes Flash v2.5 as a fast synthesis option and Multilingual v2 as a stable choice for longer material. The current model reference also documents newer V4 options. These descriptions help choose evaluation candidates but do not predict end-to-end telephone delay. A short confirmation and a long address have different requirements: one emphasizes responsiveness, while the other needs consistent pronunciation and intelligibility. Use representative text in the intended language, and check voice compatibility and current provider limits before treating a model change as a drop-in replacement.
Hold the text, voice, output format and listening conditions steady when comparing synthesis candidates. Include numbers, punctuation, abbreviations and a mixed-language proper noun. Measure first audio separately from total generation duration and listen for dropped words or awkward pacing. A faster first chunk can still produce a worse conversation if later chunks stall. Use synthetic content and authorized provider evaluation only; do not convert private call transcripts into a public benchmark dataset. Record the exact model identifier and date so later provider changes do not invalidate your conclusion silently.
Changing synthesis can affect socket lifecycle, language behavior, cancellation and billing. Lupin’s current V4 Turbo path prepares a spare connection because its streaming contract differs from multi-context endpoints. An operator should validate the actual endpoint and interruption behavior, not merely replace a model string in a saved profile. Keep the earlier configuration available for rollback and compare the same dialogue set after the change. Workspace actions still run through the configured Lupin agent backend and its permissions; selecting a new voice or synthesis model does not replace that boundary.
An evaluation report can say that the captured cascade used eleven_v4_turbo and that a separate synthetic synthesis comparison examined Flash and Multilingual candidates. It should not say the live call used all three. Preserve configuration evidence and keep provider-advertised latency separate from the intervals measured in the evaluation.
Reviewed against current checked-in Lupin routes, shared contracts and official provider references. No live calls, OAuth authorizations, app actions or purchases were performed.
Prepared with AI assistance by Lupin editorial. How we research and maintain these guides.