Voice

Two separate features, enabled separately in Agent settings.

Voice input

A microphone button next to the input. The visitor speaks, the recording is transcribed, and the text lands in the input for them to send. It is dictation — nothing is sent until they press send.

Recordings are capped at 5 MB per clip. Which model does the transcription is a platform-level setting.

Voice chat

A full spoken conversation: the visitor talks, the assistant answers aloud, and the transcript still lands in the conversation log exactly as a typed chat would.

Two engines:

• Pipeline — transcribe, answer, speak. Each turn is a round trip, so it is a beat slower, and it goes through the same retrieval and the same settings as a typed conversation.
• Realtime — a live audio session. Noticeably more natural to talk to.

Voice chat depends on a speech provider that the operator configures. Until that is in place the option stays off, whatever the agent setting says.

The microphone permission

This is the part that generates support requests, so it is worth understanding before you turn voice on.

The microphone permission belongs to the website the visitor is looking at — the address in their browser’s bar — not to Engadin. For an embedded widget that is your site. If a visitor once denied the microphone on your domain, the browser remembers, and the widget cannot ask again.

When that happens the widget links to a help page that walks them through re-enabling it for their specific browser, in English or German:

https://engadin.ai/help/microphone

Inside an iframe

An iframe gets no microphone access unless the embedding page grants it:

<iframe src="https://engadin.ai/chat?agent=your-agent&embed=1"
        allow="microphone"></iframe>

Without allow="microphone" the browser blocks it regardless of what the visitor has permitted for your domain, and the failure looks like a broken widget rather than a missing attribute.

Rate limits

Voice signalling gets a much higher per-IP allowance than chat does, because a single spoken turn fires several calls. Media calls — transcription and speech — share the same per-minute budget as uploads. See abuse limits.

When to leave it off

Voice is genuinely useful on phones and for accessibility, and genuinely noisy in an open-plan office. It also costs per minute of audio in both directions, which the cost dashboard will show you. Turning it on for a support agent is usually right; turning it on for a checkout assistant is usually not.