Voice chat

Purpose

With the voice chat, visitors talk to the assistant instead of typing: press the microphone, speak the question, hear and read the answer. That helps on a smartphone, on the move, in the workshop, and anyone who finds typing difficult. The answers come from the same knowledge as in the text chat, with the same rules.

What visitors see and hear

A microphone button next to the input field, as soon as the voice chat is switched on for the collection and the embed does not hide it.
While speaking, the transcript appears live in the input field. A short pause ends the recording and sends the question.
The answer is read aloud and shown as text at the same time, with sources as in the text chat. A tap on the stop button interrupts.
Every answer has a speaker button that reads it aloud again later.
The help text in the chat names the provider the recording is sent to.

Where

Open Edit → Voice chat in the collection's settings.

Section "Voice chat" in the collection's settings

Field	Meaning
Offer voice chat	Enables the microphone button. It appears only if the instance has set up the voice service (Chapter 26) and the chosen provider supports audio.
Audio provider	Speech recognition and voice need audio models. "Same as chat LLM" uses the collection's provider if it supports audio (Mistral, OpenAI). With a separate audio provider, the voice chat also works when the chat answers with Anthropic or Gemini.
Speech recognition (STT)	Recognition may be handled by a different provider than the voice. Gladia locks the conversation language and uses the collection's vocabulary as a recognition aid.
Speaking rate	1.00 = normal rate. A separate slider for each additional language.
Voice vocabulary	One term per line: brand and product names are recognised and corrected in the transcript. `wrong => right` hard-wires known mishearings. A separate list per language; empty = the list of the primary language.
TTS voice	The list comes live from the provider; voices you have cloned yourself appear too.

Voice and rate appear only after a speech-capable audio provider has been chosen and saved.

Data protection

The recording goes to the chosen provider for recognition; the answer is converted to speech there. The help text in the chat names the provider and its region; the collection's processing region (Chapter 20) also checks the audio providers. Recordings are not stored.

Frequently asked questions

The microphone button is missing. Three possible reasons: the voice chat is off in the collection; the embed hides "Microphone" under "Functions"; the instance's voice service is not running.

The voice speaks German with an accent. Not every provider has a German voice; some speak German with an English colouring. Try the voices in the list, or choose a different audio provider.

The chat does not understand our product name. Enter it in the voice vocabulary, with the spelling that arrives in the transcript: product name wrong => product name right.