Realtime voice chatbot

The Aimogen Pro realtime speech-to-speech chatbot built on the OpenAI Realtime API, its models, voices, shortcode and limitations.

The realtime chatbot is a genuine speech-to-speech conversation: the visitor speaks, the model hears audio directly and replies with audio, with no transcription step in between.

It is a separate system from the standard chatbot with text-to-speech bolted on, and it behaves differently.

AI Chatbot › AI Realtime Chatbot

Requirements#

Models and voices#

Voices: verse (default), alloy, ash, ballad, coral, echo, sage, shimmer.

The shortcode#

[aimogen-realtime-chat mode="toggle" model="gpt-realtime-2" voice="verse"
                       temperature="1" max_tokens="4096"
                       instructions="You are a friendly assistant for Example Ltd."
                       autostart="no" show_chat_log="show" show_user_log="show"]

Interface labels and colours are configurable too: hold_to_record, start_recording, stop_recording, not_recording, recording, record_button_color, record_text_color, record_status_color, record_status_text_color, chat_container_color, chat_log_container_color, ai_bubble_color, ai_bubble_text_color, user_bubble_color, user_bubble_text_color.

There is also an Aimogen Realtime Chatbot Gutenberg block.

Remote embedding#

The realtime chatbot has its own remote settings, remote_chat_realtime and allow_chatbot_site_realtime, working the same way as the standard remote chatbot. The domain allowlist matters just as much.

Cost#

Usage is recorded through the aiomatic_log_realtime_usage AJAX action and appears in System & Logs › Usage Logs.

Limitations#

OpenAI only. No other provider is supported for realtime.

Fewer features than the text chatbot. Personas, workflows, multi-bot and PDF chat are text-chatbot features.

Browser dependent. Microphone access and WebRTC quality vary. Test on the browsers your audience uses.

Latency depends on the network. A poor connection produces gaps and talk-over.

Transcripts are approximate. The displayed transcript is a rendering of the audio conversation, not a verbatim log.

When to use it#

Good for: accessibility, hands-free contexts, demonstrations, genuinely conversational interactions where typing is the friction.

Not good for: general website support — text chat is cheaper, more reliable, more accessible to deaf and hard-of-hearing visitors, and searchable.

Common problems#

The microphone button does nothing. The page is not HTTPS, or permission was denied. Check the browser site settings.

No audio output. Browser autoplay policy. The visitor must interact with the page first.

Session drops. Network instability, or the OpenAI session timed out.

The bill is much higher than expected. Realtime audio pricing. See above.

Related#

• Voice, speech and talking avatars — text chat with speech added
• Usage limits and credits
• OpenAI