Use browser voice safely in the website widget

What you'll achieve

Visitors know when audio starts, what the widget sends and what to do when permission, autoplay or connection policy prevents voice.

Use a current browser with microphone support and test from the exact allowed website origin.
Keep the widget frame and API origins in the site's Content Security Policy. The parent page must allow microphone access for the widget frame with Permissions-Policy.
Decide whether this widget uses Human support or AI support. Human support rings an available operator. AI support starts the configured DobroDesk voice assistant.
The widget never starts microphone capture on page load. A visitor must choose Start voice, and a blocked parent policy cannot be bypassed from inside the frame.

Open the website widget and choose Start voice. Do not treat opening the widget as consent to use the microphone.
Approve microphone access when the browser asks, then check that the widget shows an active voice state. The browser sends bounded audio frames only while the session is active.
Use Mute when you need privacy during the session. Choose End voice when you are finished. Do not close the page as a substitute for ending the session.
For Human support, wait for the operator handoff or continue with the text conversation if nobody is available. For AI support, follow the configured voice assistant prompts.

If the browser denies the microphone, choose the text conversation or change the site permission, then start voice again explicitly.
If the voice control is missing, check the exact origin, the widget status, CSP and the parent Permissions-Policy header. A policy error must be fixed on the website, not inside the widget.
If audio does not play automatically, use the visible play or continue control after a user gesture. Do not weaken autoplay policy or hide a consent prompt.
If the network or provider session drops, end the stale session, retry once and keep the text path available. Repeated failures should be recorded with the time, origin and browser.

Voice is approval- and interaction-first: no automatic call, silent microphone capture, hidden audio upload or automatic reply is allowed.