Is there a way to configure Grok Transcriber to output spoken numbers as digits rather than words?
Currently, when audio contains spoken numbers (for example, a user saying “one seven seven seven three zero three…”), the transcriber outputs spelled-out words (“one, seven seven seven…”) instead of digits (“1777 303 303”).
This makes it difficult to automatically search, format, or extract booking references, Ticket IDs, and phone numbers. We need spoken numeric data to consistently transcribe as numeric digits.
Is there a prompt setting or configuration parameter to enforce digit formatting?
If so, with the Speech-to-Text API, set format=true along with a language code (for example, language=en). This enables text formatting, which renders spoken numbers, currencies, and units as digits. It’s off by default, and format=true without language will be rejected. Full details are in the Speech to Text docs.
One caveat: long digit-by-digit strings such as booking references and phone numbers aren’t always normalized yet, and the team is aware of it. If you still see words spelled out with formatting enabled, a simple word-to-digit pass over the transcript is a reliable fallback.
But I cannot always know the language for all conversations, for all users and businesses worldwide. Scenario is like this - I’ll explain:
The requirement to specify language when setting format=true creates a major bottleneck for automated workflows. We process tens of thousands of audio recordings for an international call center handling multilingual support (English, French, Italian, Arabic, etc.). At the time of the transcription request, I do not know the language.
Since we rely on the API’s automatic language detection to identify the spoken language per call, we cannot hardcode a language parameter in advance. Requiring language for format=true effectively forces us to choose between automatic language detection or text formatting.
A few quick follow-up questions:
Are there plans to allow format=true with automatic language detection in a future API update?
In the meantime, is the official recommendation to do a two-pass API request (detect language first, then re-transcribe with format=true set) or to handle number normalization purely via post-processing code?
Ideally, the model should distinguish between general conversational quantities (e.g., spelling out ‘one million votes’) and explicit code/ID sequences (e.g., outputting a flight booking reference like ‘A12345’ or phone numbers as raw digits).