Speech-to-speech model that listens while it speaks, for noisy phone calls
The Hook
“Yeah, we can bring it in on the 5th... Not the 5th, he's talking on the phone... Oh, the 6th. Sorry, the 6th.”
Product Reveal
“The clip you just heard was entirely generated by AI. The man, the woman, the background noise, all of it. It's part of Tofu, our new speech-to-speech model that's coming in early January. 92% of people interacting with Tofu couldn't tell it was AI.”
Feature Teaser
“It's capable of generating complex soundscapes with background noise, multiple speakers, but also understanding audio as a full duplex model. That means it can hear the difference between a TV and a speaker, understand multiple speakers in a room, or even context of what you're saying. The cocktail effect has been a long problem for voice AI, and Tofu solves that.”
Feature Teaser
“Okay, um, let me go through these. So, one's already been delivered so I can't cancel that one, and the other one's already canceled. Let's see what's left. Okay, so the makeup kit's still pending, and um...”
Problem Agitation
“That was another example of our model. You can hear it thinking through a problem where it's able to in real time narrate to a user while taking complex actions on the user's behalf. What we're excited about is that this is a turning point for the entirety of voice. We believe that we've crossed the threshold into complete human naturalness...”
Call to Action
“(No spoken dialogue — subtle ambient tone)”
Synthesizable in Remotion · key components of the ai-founder-demo-suite template.
Layered audio channels display dynamic bar waveforms representing caller, agent, and ambient noise tracks.
Render stacked track rows. In Remotion, map audio amplitude data across frames to SVG path heights with interpolate() and spring damping.
A 100-cell square matrix sequentially highlights from bottom to top, locking into a solid color block matching a displayed percentage.
Generate a 10x10 array of rounded divs. Compute fill states using frame numbers and interpolate() clamped between 0 and 92.
Bullet points transition sequentially from pending dots to checkmarks with status text typing out simultaneously.
Use useCurrentFrame() conditional timestamps to stagger checklist opacity, status icons, and string slicing for typewriter captions.
Purpose. Disarm viewer skepticism by playing high-fidelity conversational audio before revealing it is synthetic.
Execution. Lower-third audio recorder player over live video.
Purpose. Anchor technical capability with an indisputable human perception percentage.
Execution. 100-node matrix chart rapid fill.
Purpose. Explain how multi-speaker audio separation works under the hood.
Execution. Multi-track horizontal waveform timeline.
Purpose. Prove real-time problem solving without awkward dead air.
Execution. Split screen live action checklist sync.
Purpose. Channel technical interest into research preview signups.
Execution. Bold logo card and interactive button cue.