Build agents

Speech recognition

In the classic pipeline, Hania transcribes what the caller says before the agent replies. The recognition engine is a single choice on the bot. All three options handle multiple languages; they differ in feel and in how much you can tune them.

This setting applies to the classic pipeline. In realtime voice the engine listens for itself, so there is no separate recognition setting. See Realtime voice.

Choosing an engine

Open the bot's Speech recognition card and pick the engine under Speech-to-text model. The tuning shown below it changes to match your choice, so you never see a control that does not apply.

EngineBest for
Flux (default)The fastest response feel. Flux notices on its own when the caller has finished speaking, works across languages with automatic detection, and lets you add custom vocabulary.
ScribeBroad language coverage with adjustable sensitivity. Choose it when you want to set exactly how quickly the agent treats a silence as the end of a turn.
Azure SpeechThe widest language coverage, including locales the other engines do not offer, Urdu among them. Hania listens in the language of the voice you picked, so an Urdu voice means Urdu recognition with nothing else to set up.

Flux options

With Flux selected, two extra fields appear:

  • Custom vocabulary (flux_keyterms). A short list of words or phrases the agent should recognize accurately, such as product names or account terms. Add them so a caller saying your brand name is transcribed correctly rather than as a near-homophone.
  • Language hints (flux_language_hints). Leave this empty to let Flux detect the language on its own. Set one or more languages when you already know what your callers speak and want recognition to lock to it.

Scribe options

Scribe adds controls for how it decides a turn is over. These are covered with ranges, defaults, and an example on the Voice tuning page.

Azure Speech and language

Azure Speech is the one to reach for when your callers speak a language the other engines do not cover. You do not set the language by hand. Hania matches recognition to the voice you chose, so selecting an Urdu voice gives you Urdu recognition automatically.