Speech-to-text

Help the AI hear callers clearly. Set how it recognizes speech and knows when to reply.

Model	Provider	Notes
`flux-general-en`	Deepgram	The default. Detects the end of a turn itself, which gives fast, natural replies in English.
`flux-general-multi`	Deepgram	The same turn detection across several languages. Use it with the language set to Auto.
`nova-3`	Deepgram	Uses voice activity detection plus a smart-turn model to decide when the caller is done.
`gpt-4o-transcribe`	OpenAI	Accurate on hard audio, but slower.

#Settings

keytermsup to 50 terms: Words it must get right: your business name, staff names, products, street names, medication names.
eot_threshold0.5 to 0.9default 0.7: How sure Flux must be that the caller has finished. Lower replies sooner but interrupts people who pause mid-sentence.
eagerbooleandefault true: Starts preparing a reply on Flux's early guess and throws it away if the caller keeps talking. It saves a few hundred milliseconds on most turns.

Tip

If callers get cut off while they think, raise eot_threshold to 0.8. If the assistant feels slow to answer, lower it to 0.6 and keep eager on.