Skip to main content

Voice input

Visitors can dictate a message instead of typing it. The chat composer has a microphone button; how it starts and stops recording depends on the Hold to record setting (see Microphone options):

  • Hold to record off (the default). Click the microphone to start speaking. While recording, a panel shows a "Listening" status and a level waveform with two controls: a ✓ button to stop and insert the transcribed text, and a ✕ button to cancel and discard it.
  • Hold to record on. Press and hold the microphone while you speak; releasing it stops recording and inserts the transcribed text.

Either way, the spoken words are transcribed into the message box.

How transcription works​

Voice input uses one of two paths, depending on the visitor's browser:

  • Browser speech recognition. On browsers that support the Web Speech API (such as Chrome and Edge), the browser handles speech recognition itself and returns the text live as the visitor talks. Note this is not on-device transcription — these browsers stream the audio to their own speech backend (for example, Google's for Chrome), not to WizChat.
  • Server transcription. When the browser has no built-in speech recognition — or if the in-browser recognizer fails partway through a session — the recorded audio is sent to WizChat's transcription service and transcribed there.
BYOI

If the chatbot uses Bring Your Own Inference (BYOI) and an endpoint is assigned the Speech-to-text role, server transcription is routed to your model instead of the platform service. See Voice input (speech-to-text) for that owner-side setup and its fail-closed behavior.

Microphone options​

Next to the microphone button is a Microphone options control that opens a small settings popover. It contains:

  • A live input-level meter so the visitor can confirm the mic is picking up sound.
  • A microphone device picker, for choosing which input device to use when more than one is available.
  • A Speech language picker (see below).
  • A Hold to record toggle, switching between push-to-talk (hold the button while speaking) and click-to-toggle recording.

These choices are remembered in the visitor's browser, so they only need to be set once per browser.

Choosing a speech language​

The Web Speech API recognizes one language per session and cannot auto-detect the spoken language — so if it is set to English, speech in another language (for example Hebrew) is transcribed as English gibberish. The Speech language picker in the microphone options lets the visitor choose the language they will speak.

  • Default. If a language has not been chosen yet, WizChat uses the browser's language — or its primary language, so a he-IL or he browser gets Hebrew — when it matches a supported option, and otherwise falls back to English (US).
  • Browser speech recognition. The chosen language is used as the recognition language, so speech is transcribed in that language.
  • Server transcription. The chosen language is passed to the transcription service as a hint only when it has been explicitly chosen; otherwise the service is left to auto-detect the language, so nothing changes for visitors who never open the picker.

The picker offers the following languages, listed alphabetically:

Arabic, Chinese, Dutch, English (UK), English (US), French, German, Hebrew, Hindi, Italian, Japanese, Korean, Polish, Portuguese, Russian, Spanish, Turkish.

note

Which languages a browser can actually recognize with the Web Speech API varies by browser and platform. The list above is the set of languages WizChat offers in the picker.