Skip to main content
Voice agents often receive spoken-form transcripts while tools and APIs expect structured written values.

​
Where ITN fits

ITN converts the transcript before an agent extracts values or calls a tool. This reduces avoidable formatting failures at the API boundary. For example, an order lookup may use an exact string key. If the transcript contains seven eight three two nine but the lookup receives the spoken words instead of 78329, the tool may not find the order. The same issue applies to phone digits, confirmation codes, and amounts.

​
Values that voice agents exchange

Common values include order IDs, phone numbers, dates, times, amounts, room codes, email addresses, URLs, and alphanumeric identifiers. See the complete supported forms.

​
Ambiguity affects tool calls

two thirty can be a time, a price, or an identifier. A format that is valid in one sentence can be wrong in another. If the wrong interpretation reaches a tool call, lookup and transaction failures follow. Premove ITN generates valid alternatives, scores them in sentence context, and uses exact decoding to select compatible edits. The frozen benchmark measures this behavior on synthetic voice-agent stress cases. It is not a production- traffic accuracy claim.

​
Put ITN at the right boundary

Normalize transcript text before converting it to a tool argument. Then apply the tool’s own validation rules. ITN can reduce formatting ambiguity, but it does not confirm that the ASR transcript is correct or that the user authorized a transaction. Continue with Getting started, review supported forms, or inspect the benchmark method and limits.