Cooking apps (recipe apps, guided-cooking apps, smart-oven companions) run on timers, and timers do not know your stove, your pan or how wet your onions are. Smart thermometers measure the inside of a roast, not what is happening in an open pan. Sizzle judges doneness from the sound of the pan itself: the hiss of water leaving, seeds popping, a crackle dropping in pitch. That is the cue every experienced cook already listens for, now said out loud.
The data it runs on
The phone microphone: a short rolling window of pan sound, turned into a spectrum on the phone. Raw audio never leaves the device unless the cook opts in to help training.
An audio-capable model: every few seconds a compact feature summary (sizzle loudness by band, pop rate, trend) is classified into states. Early versions send a 5 second clip to an audio-input model (Gemini Flash or GPT audio models via OpenRouter); later, a small on-device classifier trained on opted-in clips.
Limits and gaps: extractor fans, radios, boiling pots nearby and very quiet induction hobs all blur the signal. Sound cannot see colour, and it cannot tell how cooked the inside of a steak is.
Fallback: when the signal is weak the app says so and falls back to a timer adjusted for your stove, plus the "confirm by eye" photo check with a vision model.
1. The mission
Give every home cook the ears of a line cook. Sizzle listens to the pan so you can chop, talk and look after children, and it tells you the moment to stir, flip or pour.
2. Why people can't resist it
Magic that is real. The phone says "Flip now" a second before you would have guessed. That moment is the screenshot.
Hands-free relief. No sticky screens, no timer resets. It talks; you cook.
It gets to know your stove. Each session teaches it how hot your burner really runs, so it gets better the more you cook.
3. How it works for real
Calibrate: 3 seconds of room sound; steady hums such as fans are measured and filtered on the phone.
Listen: Web Audio or native audio APIs compute a 40-band spectrum 20 times a second on device.
Classify: every 5 seconds, features (and in v1 a short clip) go to an audio-capable model (Gemini Flash via OpenRouter) that returns a state: heating, sizzle, moisture popping, quieting, browning, done.
Cue: a small rules engine per dish maps states and timing to spoken cues, read by the phone's own text to speech, so speech costs nothing.
Confirm by eye: a photo goes to a vision model (Gemini Flash or Claude) that compares colour against a doneness scale.
Learn: the "how did it turn out" tap updates a per-burner offset, stored on the phone.
4. Cost per use
About $0.02 per cooking session (roughly 60 short classification calls at a fraction of a cent each, plus one optional photo check). On-device classification later brings this close to zero.
Free: onions, rice and 3 other everyday dishes, unlimited sessions. Sizzle Plus ($3 a month): the full library (searing, frying, tempering, caramel, deep frying with temperature alerts), multiple stoves and household sharing.
5. Growth loop
The "called it by ear" card after each session is a little brag that people post. Friends who cook the same dish get a shared link that starts the same cue plan. Recipe creators can publish "listening recipes" with Sizzle cues built in.
6. Risks and guardrails
Food safety: Sizzle never decides that meat is safe. It always asks for a thermometer on meat and says so out loud.
Fire safety: it is not a fire alarm and says so; hot oil sessions have a hard stop timer and a "never leave the pan" reminder.
Privacy: the mic is only on during a session, with a visible live indicator; audio is processed on the phone and summary features are sent, not conversations.
Honesty: when the signal is unclear it says "I can't hear the pan well" instead of guessing.
7. MVP build order
Web Audio spectrum view with a live mic, plus room calibration.
One dish (onions to golden) with hand-tuned rules from loudness and pop rate.
Gemini Flash audio classification every 5 seconds, compared against the rules.
Spoken cues with speechSynthesis, and the confirm-by-eye photo check.
The result feedback tap and per-burner learning, then steak, plantain and tadka.