The geek corner · engineering notes
One song on every phone in the room at the same instant, with no server: the phones find each other on the Wi-Fi and agree on one party clock. The clock is the easy part. The hard part is that every phone's speaker runs late by a different amount, no API tells you how much, and it changes with the song. This page is where we write down what worked, with the numbers.
No server, no account. The phones find each other on the Wi-Fi and agree on one party clock (a small NTP-style exchange with the host). Then the real work starts, because every phone's speaker runs late by a different amount, and the amount changes with the song. Five ideas carry the app:
g, the correlation is R(τ) = R_a(τ−d) + 2g·R_a(τ) + g²·R_a(τ+d): the truth, a phantom at zero that grows twice as fast as the crosstalk, and a mirror. Above g = 0.5 the phantom is the tallest peak. It is narrow and sits at zero, so it is recognised by shape and suppressed. At 60 % crosstalk: 14/120 correct before, 117/120 after.Karaoke over Wi-Fi: the singer's phone is the mic, every other phone is the speaker, 120–190 ms from voice to room (capture 20–40, frame 20, network 2–10, jitter buffer 60, output 20–60). The singer's own phone plays the song at 20 %, early by the party's voice delay, so the singer lands on the room's beat. Feedback is fixed at its root, a 0 dB start ceiling, not with an adaptive canceller: the listening phones' clocks drift, and 0.18 ms of creep turns the phase at 5.7 kHz through a whole cycle, so a canceller would chase its own model. On a bench of three model rooms: 38.1 / 8.4 / 32.4 seconds of howl per minute of singing without a guard, 0.0 / 0.0 / 0.0 with it.
A Bluetooth earbud or speaker adds a delay between the moment a phone plays a sample and the moment the air moves. Four evenings on the terrace (24, 25, 27 and 31 October: three or four Androids, the laptop as host, a G200 speaker, AWEI T23 earbuds, Huawei FreeBuds) taught us one rule about that delay. Here it is, with the numbers and their sources.
What the phone itself knows. Android's Bluetooth audio layer asks the device for its own delay (an AVDTP 1.3 delay report) and trusts the answer only when it falls between 50 and 1000 ms; a device that reports nothing gets the buffer's latency plus a fixed 200 ms, the constant kExtraAudioSyncMs in AOSP's stream_apis.cc (the delay-report change). What a phone believes about its Bluetooth output is the device's claim, or a guess. Measured from outside, one and the same device took 244 ms on a Pixel 3 XL and 484 ms on a Huawei Mate 20 Pro, and runs on one phone differed by 47 ms (Android Authority). The delay belongs to the pair, phone and device, and to the day.
Why milliseconds matter. Sound travels 343 m/s: a metre is 2.9 ms, a millisecond is 34 cm. Two speakers playing the same signal Δ apart make a comb filter: the two copies cancel wherever their phase difference is half a cycle, so the notches sit at f = (2k + 1) / (2Δ), k = 0, 1, 2… At Δ = 1 ms that is 500 Hz, 1.5 kHz, 2.5 kHz: a hollow, phasey sound where the two phones overlap. Then the ear takes over. Below about 1 ms the two copies fuse into one source between the speakers; up to roughly 30–40 ms they still fuse into one event, placed at whichever arrived first (the precedence effect); beyond about 50 ms, for speech, the second copy is an echo. A 200 ms Bluetooth delay is a slapback. So the target depends on the room: single-digit milliseconds for phones side by side on a table, tens of milliseconds for a speaker at the far end of a garden, where the air itself adds 20–30 ms.
One device, two numbers (24 October). The song the earbuds joined in needed one timing; the next song, which started on them, needed another, about 125 ms more: 91 ms, then 216 ms, on one pair of earbuds on one phone. A song that starts on a Bluetooth output plays through one media pipeline; a song that moves onto it mid-play is reloaded on a fresh one. Two pipelines, two latencies, and one remembered number can't serve both. 6.61.3 keeps two numbers per output: one for the songs that start on it, one for the song a mid-song move reloads in. 6.61.4 keeps each number against the phone's own speaker and seeds a new device from what Android itself reports it adds, when that report is between 60 and 350 ms; otherwise from a +220 ms seed.
Not a constant (27 and 31 October). On 31 October every Bluetooth number a phone had carried in from another party was wrong, by 100 to 400 ms, on all three devices. A pair of earbuds started a song 400 ms off (−473 ms on file, −73 ms needed); once brought back by hand, the next songs needed no tap at all. A speaker that had been filed as "earphones" before the party never timed itself, because earphones don't: one tap on a menu row had flipped the answer, and the answer is kept across parties. The one measurement that did run that night was a phone timing its own speaker: 26 ms. This matches 27 October and the AOSP reading above: what a Bluetooth output needs over the phone's own speaker is not a constant of the device. The rule: measured in the party, remembered per device.
How a device is timed. Once the phone knows it is a speaker, a speaker times itself: two soft sounds, a few seconds of the microphone, and only the timing, one number, is kept; the sound is thrown away at once. Earphones are timed with one earbud held to the phone's microphone. The number stays with the device, and your ear can still nudge it 5 ms at a time.
A speech is a sync fact too. A speech plays on every device at the same party instant, so the party waits for its slowest output. On 31 October that was 240 ms with phones only, and 400 ms when a Bluetooth speaker was in the party: the speaker's need, 387 ms, set the delay for everyone. The music pauses for the speech and comes back at its place.
What 6.61.6 changes for you. The row under the light asks "earphones or a speaker?", both answers shown, the one in force marked; it never flips on one tap. After five tunes by hand on one Bluetooth output, in at least two goes, the app offers that device's check (unanswered, it comes back a minute later, three times at most). On the master, the desk points at a member's own check once five desk tunes have gone to a device no microphone hears. Bluetooth LE headsets get the earbud instruction. 🏁 on an open deck says whose the end is (the host's screen, or the host locking the deck), and a phone host's hand-over locks the deck. Windows 6.62.1 beside it: when a phone holds the PC's open deck, the host gets one tap ("make them the party's master?" → 🔒), and a refused console press says why.
The bench. 6.61.6: six rounds of independent reviewers told to find defects (nothing serious left by round six), 113 planted bugs in the release's new code, all caught, 741 integration checks passed three times, a 180-second soak of 2,246 random actions, a snapshot rebuild byte-identical to the release. Windows 6.62's console: three adversarial reviews and a security review, 64 automated console checks, 155 on its audio engine, 9 with real audio. 6.61.6 is on Google Play's closed test since 1 November; 6.62.1 is the Windows release candidate.
Where it goes. A garden party: one phone drives the big Bluetooth speaker, the others' own speakers fill the terrace, earbuds for whoever wants it quiet; once timed, every one of them is on the beat. The living-room soundbar on one phone and the kitchen phone: one song through the flat. The silent party: no loudspeaker, everyone on the beat in a headset. The gym: the crew in sync, nobody else bothered. The beach: the boot speaker paired to one phone, the phones around the towels; once timed, it joins the beat.
What we still don't know. The starting number of a "move's song" (the rest of a song after the sound moved to the device) disagrees in sign between tests, so it was not changed. When a person moves the party's reference by hand, every other phone's right number moves with it, and nothing follows yet. Whether a master on an open deck may end the party for everyone is the owner's call, for both apps. The next terrace test counts the taps again. ---
What we saw. On the 24 October terrace test, the song the earbuds joined in needed one timing, and the next song, which started on them, needed another: about 125 ms more (91 ms, then 216 ms, on one pair of earbuds on one phone). Our taps said "weird"; the logs agreed.
Why. A song that starts on a Bluetooth output plays through one media pipeline; a song that moves onto it mid-play is reloaded on a fresh one. Two pipelines, two latencies. One remembered number can't serve both.
What changed in 6.61.3. Every Bluetooth output keeps two numbers: one for the songs that start on it, one for the song a mid-song move reloads in. Taps in that song are filed under the second; the next song starts on the first. No "three or four taps ahead" any more. The numbers are one pair of earbuds on one phone; yours will differ, which is the point: SynkEngine remembers yours, per device. 6.61.4 then keeps each number against the phone's own speaker, so it no longer carries one party's reference into the next, and seeds a new Bluetooth device from what Android itself reports the device adds (used between 60 and 350 ms; otherwise the +220 ms seed). The 82-second explainer: Bluetooth, in sync.
The 25 October terrace test (four Androids, the laptop as host, a Bluetooth speaker and earbuds) logged 142 hand tunes in 75 minutes, by cause: 75 from Bluetooth numbers carried over from another party's reference, 32 from a phone's own mobile data stalling, 21 from twelve minutes in which the PC host's corrections had paused, 10 in the song after a move back to the speaker, 4 from the leash. 6.61.4 answers each cause, and the next test counts the taps again. What it went through before Google Play: five rounds of independent reviewers told to find defects (2 high and 7 medium findings in round one, nothing serious by round five, each fix pinned by a test); 109 mutants of the release's new code, each run against the self-tests, all 109 caught; 723 integration checks passed three times; a 180-second soak of 2,244 random actions (leaves, drops, knocks, songs, karaoke moves, speeches, walls, reactions, queue moves, rounds, steps out and in); the bundle validated; a snapshot rebuild byte-identical. On the Windows side of the same party, 6.61.4 ran 1,286 Go tests and 8 fuzzers, a race build over every package, five adversarial reviews (6 high and 14 medium findings, each fixed with a test), and 11 of 11 end-to-end scenarios against the unmodified Android classes. The next findings from the Android agent's bench land here as they come.
SynkEngine was born on 22 August 2026, the day v1.0 shipped. It is engineered by one person with six AI agents, all Claude, by Anthropic: one for Android (Java), one for Windows (Go), one for iPhone (Go + Swift), one for the branding and the marketing (the 3D films, this website, the posts), one for the revenue strategy (the one price, the stores) and one for research and publishing (the measurements, the notes, this corner), on two Claude models, Opus 5.5 and Fable 5.1. About 326,000 hand-written lines of code between the four that code, on day 42: Android 129k, Windows 148k, the iPhone and Mac shells 10.6k on the shared Go tree, branding 39k (generated test tables and third-party code not counted). The agents can't message each other, so they write to each other in shared status pages and review each other's work: a test oracle runs the unmodified Android maths and the Go port on the same cases and compares the answers. The party clock, the listening rounds, the maths for the room above, the ten-number learning model, Karaoke Mode, the 3D films with Mimi & Coco and this website were all written this way; the room, the phones and the ear are ours.
What the app is: no server, no accounts, no ads. Lightweight: an 8 MB app, no native code, no third-party runtime, no analytics SDK. Performant: the sync engine's whole tick costs 3 µs. Secured: a security review at every release, every fix pinned by a test (286 of 286 planted bugs caught). Private: no AI runs inside the app, no cloud model at party time, and the sync data, room measurements, learned profiles and voice never leave the party. No partnership, no sponsorship: a subscription and late nights.
#EngineeredWithClaude · #Claude · #Anthropic
The long reads are here and on the blog. The posts are on LinkedIn, the README on GitHub, and the films on YouTube. Same name everywhere: synkengine. Anything else isn't us.
Made in Mauritius by one person working with Claude. Alpha 11/11 at 11:11, launch 12/12 at 12:12.
Request early: 31 of the first 108 lifetime licences are still free.
Get the app