Settings

Saved in this browser.

Microphone

DeviceInput channel

For audio interfaces with two inputs (Scarlett, etc.). Auto picks the input where the voice is louder.

I sing with headphones (echo cancellation off — honest sound)

Audio latency

It takes tens of milliseconds for the track to reach your ears and your voice to reach the analysis. Without compensation it looks like you are late. The browser estimates the latency itself (now ≈ 30 ms), but measuring it is more accurate.

Automatic

Calibration (10 seconds). You will hear 8 clicks. If you sing with headphones, hold an earpiece to the microphone. If the clicks come from speakers, do nothing. You can also clap exactly in time with the clicks, but that is less accurate.

Recorded your voice and it lags behind the track in the studio? The studio editor has Voice and track together: shift the voice by ear or match it to the original notes and press Apply to future recordings — the latency will be saved here.

Notes and scoring

Note names

Tuning (A4)Usually 440 Hz. For a song, the tuning is detected from the track.Default strictness

Within the tolerance a note counts fully. Slightly off counts too, but for fewer points: a semitone off gives half.

Game

Note grades, combos, confetti and award notificationsAward sounds

XP, levels and achievements are always counted — only the effects are turned off. All rewards are in Awards.

Separation server (GPU)

Makes the hardest case cleaner — an original without an instrumental: a neural network separates the voice from the music and CREPE detects the notes. Without the server the voice is removed right in the browser (quick mode: approximate notes, a faint echo of the voice in the instrumental). The server runs on a computer with a GPU; everything else works without it.

Not found

Start: project folder → server\Запустить сервер.bat. For the AI coach, put a Claude API key in server\.env (line ANTHROPIC_API_KEY=…). The key stays on the server and never reaches the browser.