AssemblyAI format_text false: 'one hundred' vs 100 and Sume STT

On August 20 AssemblyAI added spoken-form output for async universal-3-5-pro: set format_text: false and a speaker saying "one hundred" comes back as "one hundred", not "100". Sume STT has no format_text field or equivalent toggle, so you cannot choose between formatted and spoken form on the request.

AssemblyAI's behavior is from its changelog; Sume's from the /v1/stt-1.0/transcribe schema in the API reference, read 2026-10-01. The Sume docs do not say how numbers are written in the output.

What did AssemblyAI change?

The changelog says format_text: false "preserves the transcript in spoken form rather than formatted text: numbers, dates, and other entities are left unnormalized". It calls this useful for verbatim workflows such as medical dictation, and says it is available for all accounts.

What can I set on a Sume STT request?

audio_url, optional segmentation, language_code, duration_seconds, metadata and the job mode fields. The schema is closed, and provider knobs are fixed server-side. Nothing controls number or date normalization.

Can I get verbatim text another way on Sume?

Only by testing. Run a clip that contains numbers, look at the words[] tokens, and see whether they are digits or words. The docs promise neither. If you need guaranteed spoken form, for example for medical dictation, pick a service that documents that switch.

What if I need digits from spoken words?

That is the easy direction: normalize in your own code from the tokens you receive. Each token keeps start and end, so the times survive your edit. The related Deepgram numerals post covers phone digits.