Two switches

Mic input

Turn on Voice Input and a microphone button shows in the widget. The visitor speaks, the audio is turned into text with Whisper, and that text is handled like a typed message. The recording is discarded once the text comes back. Groq's Whisper is used when it is available, because it is faster. Otherwise the audio goes to OpenAI Whisper.

Spoken replies

Turn on Voice Reply and the answer can be read aloud with OpenAI text-to-speech. If the visitor asked with the mic, the reply is spoken back. There is no extra "auto-play" setting for that.

When the mic helps, and when speech helps

You do not need both. A lot of sites only want the mic.

The mic is useful when:

• Most visitors are on a phone. Tapping the mic is faster than typing a long question.
• Typing is hard for the person, because of a motor problem, dyslexia, or because they would rather talk.

Spoken replies are useful when:

• The person cannot look at the screen, such as a reception tablet they are standing in front of.
• They cannot read the screen easily.
• The answer is short. A couple of sentences sounds fine. A long policy read-aloud does not.

The seven voices

There are seven voices. The chat header shows a first name for the one you pick. The picker itself only labels them Neutral, Male, or Female. There is no preview button. Try one in the agent's test chat.

Name in the chat	Voice	Label in settings
Alex	Alloy	Neutral
Ashley	Ash	Male
Blake	Ballad	Male
Ethan	Echo	Male
Sage	Sage	Female
Sophie	Shimmer	Female
Victor	Verse	Male

Speed

The slider runs from 50% to 200%, in steps of 10. 100% is normal speed. The number on the slider updates as you move it. The audio does not. Check the result in test chat. Slower is easier to follow for a long answer. 100% is fine for a short one.

What spoken replies cost

A spoken reply is an extra speech request on top of the normal chat. Longer answers cost more to speak. The same text, voice, and speed is saved for 7 days. If 50 people ask "What are your hours?" and the answer text matches, the speech is generated once and reused.

On a longer answer, the first sentence can start playing while the rest is still being written. A line in the system instruction such as "Keep spoken answers to two or three sentences" keeps both the chat and the speech shorter.

The mic notice

The browser still asks for microphone permission. The first time someone taps the mic, ChatNexus also shows a short notice: the audio is sent to be transcribed and is not stored. The checkbox says "Don't show again." Recording can start when the browser allows the mic. They do not have to tick that box before speaking. If they do tick it, the notice stays hidden for that agent in that browser.

If the host page blocks the microphone inside the embedded frame, ChatNexus opens a small popup on chatnexus.cloud so the browser can ask for permission there. See how to fix microphone permission in an embedded chatbot for the details. Repeat spoken answers are cached, which keeps voice costs down.

Where the switches are

On the agent, open Voice Settings:

• Voice Input — shows the mic. Every plan. Free is capped at about 10 minutes a month.
• Voice Reply — reads answers aloud. The screen marks this as Starter or above.
• Voice — one of the seven names above.
• Speed — 50% to 200%.