Listening tool

Sonogram

Ultrasound for songs.

Drop in a song you know. Listen to just its center. Listen to just its sides. Loop the chorus. Watch what falls out when one part disappears. Nothing leaves your device.

Drop an audio file here — or click to choose

MP3, WAV, M4A, FLAC, OGG, AAC, OPUS, AIFF. Stereo files unlock center / sides modes; mono files will play full-mix only.

Files greyed out in the picker? They may be Apple Music or Spotify downloads — those are DRM-protected and no browser tool can decode them. Pick MP3, WAV, or M4A files from your own library instead (in the Files / Finder view, not the Music app).

Runs entirely in your browser · nothing is uploaded · nothing is stored

Now listening to

—

Drag on the waveform to loop a section — a chorus, a bridge, a two-second sigh. Click anywhere outside your selection to reset to the whole song.

0:00 / 0:00 Loop selection Volume

What plays

Full mix As the record was released. Center only What sits dead-center: usually lead vocal, kick, snare, bass. Sides only (vocal remover) What was panned outward: pads, doubles, reverb tails, most guitars. The centered vocal disappears.

Full stem separation

AI · slow · optional

Real per-instrument separation into four tracks: vocals, drums, bass, and everything else. Much cleaner than the center/sides trick above — but it costs a one-time ~172 MB model download (cached after that) and then a few minutes per song. The center/sides controls stay usable throughout; this is a longer-form pass, not a replacement.

Listen for a device

A pre-cue and a channel choice, together. From the mourning-archetype week . Tap one to pre-set the ear.

What this tool actually does

Center and Sides are the classic mid/side trick — the same math old karaoke machines used to strip vocals. It works because most pop mixes put the lead vocal dead center. It doesn't work when the vocal is panned, doubled to the sides, or drenched in stereo reverb — you'll hear a ghost of it in Sides mode. That failure is teaching content too: it tells you the engineer decided this vocal wasn't supposed to sit alone.

Full stem separation is the longer-form pass above. It runs a Demucs neural network (HT-Demucs v4, ~172 MB) locally in your browser via WebGPU or WASM — still zero uploads, but a real few-minute wait per song. The stems it produces are cleaner but not perfect: complex arrangements can leak a little between tracks, and instruments the model doesn't recognize (harpsichord, sitar, unusual synths) usually end up in other. That imperfection is worth naming to the room the first time it appears; students trust a tool that admits what it can't do.