How to disable/control ASR "smart formatting" that inserts slashes into dictated ID numbers (Chilean RUT)?

Support Help
1

I’m running a voice agent (agent_id: agent_ab5d6f0301c136cf47aef4a4fd) that asks callers to dictate their Chilean national ID number (RUT), a string of digits plus a check character (e.g., spoken as “doce millones veintiocho mil ochocientos siete, guion K” or similar, meaning 12028807-K).

The real-time transcript consistently comes back with the digits auto-formatted with slashes, as if the ASR is recognizing the number sequence as a date-like entity, e.g.:

User: “Mi RUT es el 12/028/807-K.”

The caller never says “slash” — they only dictate numbers. This looks like automatic entity/number formatting happening at the ASR layer (Deepgram/Azure/Soniox, auto-selected by Retell based on language/optimization settings).

Reference call for reproduction: call_c198b8e402b313867b9da39cc0a (agent_version 37, dynamic variable rut_dictado).

Questions:

Is there a setting (documented or not) to disable this kind of automatic number/date formatting in the raw ASR transcript, independent of Transcription Mode (speed vs. accuracy) or Boosted Keywords?

Does the ASR provider chosen for a call (Deepgram/Azure/Soniox) affect how aggressively this smart-formatting is applied? Is there a way to influence/force provider selection to reduce this?

Is there a way to access a “raw”/unformatted transcript stream (pre entity-formatting) for use in extract-dynamic-variable or code nodes, so downstream logic doesn’t have to strip formatting characters?

I already work around this downstream (normalizing the captured value into a clean RUT format before validation), so this isn’t blocking, but it’d help to understand if there’s a supported way to prevent it at the source, especially for other numeric identifiers (phone numbers, account numbers) where the auto-formatting could cause similar issues.

2

Hey @rcarcamoc I’m checking your Call ID with the team and will get back to you as soon as I have an update.

Thank You

3

Hey @rcarcamoc Short answer: There is no Retell setting today that disables ASR number/entity formatting independently of Transcription Mode or Boosted Keywords.

We are investigating this and will get back to you once we have better understanding.

Thank You

4

Hey @rcarcamoc We looked into call_c198b8e402b313867b9da39cc0a.
The slashes come from the speech-to-text provider, not from your agent’s logic. On this call the caller read the RUT in groups (“trece, cero veintiocho, ochocientos nueve, guion K”). In Accurate mode for Spanish, the transcriber writes grouped numbers like that with separators, so the result looked like a date. The extract-dynamic-variable node then copied the value exactly as written.

To your questions:
1: There is currently no setting to turn off number formatting in the transcript.
2: Yes, the provider affects this. You can choose one with stt_mode set to “custom” and a custom_stt_config (provider plus endpointing_ms). For example, Deepgram does not add these separators, but it often returns the digits as spelled-out words (“trece cero veintiocho…”), so your extraction would need to convert them. We suggest testing it before you roll it out.
3: There is no raw, unformatted transcript stream available to extraction or code nodes today.
The most reliable fix is the one you already have: normalize the value after you capture it. You can also tighten the rut_dictado variable description, for example: “Return only the digits, a hyphen, and the check digit (e.g. 12345678-K), with no dots, slashes or spaces.” That way the extraction step removes the separators before your validation runs.

Thank You