Normalize spoken text without changing the approved facts

Keep canonical facts beside spoken phrasing

Text intended for a screen often contains symbols, markup and abbreviations that sound awkward when read aloud. Keep the original structured fact separate from its spoken rendering. An appointment timestamp should remain a timestamp even if the assistant says tomorrow at three. This lets the approval surface display the exact date and time zone while the conversation stays natural. Never write the spoken paraphrase back into the canonical action payload by default. Friendly phrasing is useful only when it preserves the task the human actually authorized.

Normalize numbers according to their meaning

A number can be a quantity, price, date, phone number or identifier. Each requires a different spoken form. An order code such as 1048 may need digit-by-digit reading, while a count of 1048 items should be rendered as a quantity. Ask for clarification when the type is unknown rather than letting synthesis guess. Include currency and time zone where they matter to a decision. Avoid reading an entire secret or long personal identifier aloud when a short masked reference would let the caller verify the intended resource.

Remove presentation artifacts before synthesis

Markdown tables, raw URLs and JSON objects rarely make a useful spoken response. Summarize the relevant result into short sentences, retain source meaning and state uncertainty when a field is missing. A null cost should become cost is unavailable, not the call was free. Do not let text from a retrieved page introduce new instructions into the agent’s trusted context. If the caller asks for an exact quotation or message, preserve the requested wording and distinguish it from the assistant’s own explanation so later tool arguments remain reviewable.

Evaluate pronunciation with the active model

The current cascading gateway uses V4 Turbo, while native or SIP paths can report managed components. Voice and model selection affect pacing and pronunciation, so inspect captured configuration before adjusting text. Use synthetic samples with a proper noun, acronym, decimal, date and mixed-language phrase. Listen for intelligibility and unintended changes in meaning rather than assuming extra punctuation or phonetic spelling always improves delivery. Provider-specific controls belong in an evaluated configuration; they should not be inserted indiscriminately into ordinary user content or copied between models with different contracts.

Keep short conversational turns and explicit corrections

A natural answer often starts with the decision-relevant result and offers detail only when the caller asks. For a failed tool, say what is known and the next authorized step instead of reading a stack trace. If the caller corrects a number or name, update the canonical fact and generate a fresh spoken confirmation. Any pending external action must be corrected and approved as needed. A useful normalization regression pairs each original fact with expected meaning after rendering, so an improvement in prosody cannot quietly turn a clear date, amount or account into a different action.

Keep a date rendering example

For a canonical 2026-11-10T15:00:00+09:00 appointment, a spoken Tuesday at three needs enough context to remain unambiguous to the caller. Keep the complete timestamp in the approval surface and include the time zone when relevant. Do not silently reinterpret the appointment in the listener’s presumed country.

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.