AI-generated Answer for al_61528
Short answer: There isn’t a documented STT/Agents setting to force “digits-only” transcripts or to disable punctuation/ITN. Scribe’s API doesn’t expose a numeric-format toggle, so you need to constrain the LLM behavior and add validation on extraction. (elevenlabs.io)
What to do
-
Tighten your system prompt for number capture
- Instruct the agent to ask users to speak digits one-by-one, and to repeat back exactly the captured digits with no separators. The prompting guide explicitly recommends formatting numbers for clarity in voice (e.g., pausing between groups) — adapt that for collection and confirmation. (elevenlabs.io)
- If your audience may speak Spanish (e.g., “28016” is a Madrid postcode), add Spanish to the agent and enable the Language Detection tool so the agent operates in the user’s language, which can reduce mis-hearings of numbers. (elevenlabs.io)
-
Use Data Collection with explicit formatting rules
- For each field, set a description that states the exact expected format and rejection criteria, e.g.:
- phone_number_9: “Extract a 9-digit phone number, digits only (no spaces or symbols). If not exactly 9 digits, return null.”
- postcode_5: “Extract a 5-digit postcode, digits only. If not exactly 5 digits, return null.”
- The Data Collection docs recommend including format requirements and examples to stabilize extraction. (elevenlabs.io)
-
Add server-side validation in your webhook
- In your post-call transcription webhook, validate with strict regex before accepting values (e.g., ^[0-9]{9}$ and ^[0-9]{5}$). The post-call webhook delivers the full transcript and extracted data so you can enforce these checks and treat failures as nulls/retry prompts. (elevenlabs.io)
-
Regression-test with Agent Testing
- Create Tool Call Tests that validate parameters with regex so you catch prompt regressions that start introducing separators (example in docs shows regex validation). While shown for tool parameters, the same patterns help you verify your prompts/LLM reliably produce digits-only outputs. (elevenlabs.io)
Why the hyphen shows up
- The LLM can “normalize” numbers or add separators when composing text; similarly, STT can insert punctuation. ElevenLabs documentation advises controlling numeric rendering through prompts/normalization strategies when working with spoken numbers. Apply that guidance here but in reverse: instruct “no separators” for capture, then confirm back exactly what was heard. (elevenlabs.io)
If you need references to confirm there’s no STT toggle: see the Speech-to-Text capability page and the Create
foggy venture
gentle cairn