Full-duplex audio synthesis model for natural conversation with AI characters
The Hook
“(No spoken dialogue — ambient electronic synth chime)”
Feature Teaser
“I've been researching full-duplex lately, you know. Researching full-duplex? That sounds interesting. What exactly are you looking into? It's called full-duplex dialogue, or two-way dialogue, but— Yeah. With duplex, like, the other person can give little responses while you're talking— Yeah. Oh, right, like how you just said 'yeah'—you can chime in like that, or say 'mm.' It's a spoken dialogue model that does that. Existing turn-based models couldn't do that, though.”
Product Reveal
“On days when I'm this tired, how should I spend the evening? Have a warm drink and do some light stretching. Mm-hm. I see. Stretching, huh? Stretching sounds good. Any stretches you recommend? Just slowly rolling your neck and shoulders can help you relax. Mm-hm.”
Product Reveal
“My mom's birthday is next week. What do you think I should get her? Hmm, maybe start by thinking about her hobbies and the things she likes. For example— Oh, wait, wait. My budget's about 3,000 yen, and my mom isn't really into flowers. Got it. Around 3,000 yen, and no flowers.”
Feature Teaser
“Hey, on a night when you're this tired, how do you think you should spend it? Hmm, maybe start with a warm drink and give your body a rest. Like hot milk... I see. Hot milk. Hot milk sounds nice. That'd be so relaxing.”
Product Reveal
“Shizuku, could you check the recent average gas price nationwide? Got it. I'm looking it up right now, so just give me a sec. Thanks for waiting. It looks like it's about 170 yen per liter on average nationwide. It's gone up quite a bit, hasn't it?”
Call to Action
“(No spoken dialogue — gentle water droplet synth note)”
Synthesizable in Remotion · key components of the duplex-voice-telemetry template.
Dual parallel waveform tracks plotting user and avatar audio with overlap highlights
Render parallel SVG bar charts animated with useCurrentFrame() and Math.sin, changing bar color to red when both exceed threshold.
Small pill tags indicating conversational intent states like 'Both talking' and 'Stops'
Use spring() scale-in on state change with conditional rendering based on dialogue timestamps.
Continuous subtle vertical breathing float with mouth and eye animations
Apply interpolate(Math.sin(frame / 15), [-1, 1], [-4, 4]) to avatar container translateY.
Purpose. Establish conversational latency elimination via backchannel responses
Execution. Split view of character and dual waveform telemetry monitors
Purpose. Showcase natural flow without awkward turn-end silences
Execution. Smartphone frame with synchronized animated speech captions
Purpose. Prove instant speech abortion upon user barge-in
Execution. Waveform cut indicator with explicit visual 'Stops' callout
Purpose. Demonstrate multitasking tool usage without speech freezes
Execution. Tool execution pill animations embedded directly along the audio track