Learn, explore, repeat !
every song becomes a playground

Fine stem separation — up to 17 stems

A new separation model, mvsep_mega_fine, runs a three-stage pipeline: Demucs htdemucs_6s for the coarse split, then DrumSep (inagoy, HDemucs) splitting the drum kit into kick, snare, toms and cymbals (hi-hat folded into cymbals), then MVSep Mega BS-RoFormer (ZFTurbo) heads splitting the remaining stems with Wiener masks.

• Up to 17 stems — lead and backing vocals, drums plus kick/snare/toms/cymbals, bass, electric and acoustic guitar, piano, organ, synth, brass, winds, strings, and other.
• Roughly 1-2 minutes per song on a GPU.
• Model weights (~1.5 GB) download automatically on first use.

Requires an NVIDIA CUDA GPU with about 4.5 GB of usable VRAM. This model does not run on CPU — the other separation models still do.

Mixer improvements

• Chords view is now a stage chart — a grid of four bars per system, lyrics under every beat, chord names that scale to their cell, a clearer "now" cue, and auto-scroll that keeps the active system near the top for maximum look-ahead.
• Audible scrubbing — dragging the timeline ruler moves the playhead and plays short slices, so you can find a passage by ear.
• Richer loop controls — draggable edge handles, fields accepting a timecode (1:23.45) or a bar (b17), a clear button, and a shared snap-to-beat toggle. Defining a loop is Shift+drag, on the ruler or on a waveform.
• Lyrics popup became a stage prompter — real font scaling up to 3x with reflow, size remembered between sessions, active line parked mid-screen.
• Faster loading — tracks and server-side waveforms appear immediately, and each stem decodes as its own download finishes.

Lyrics now come from LRCLIB, aligned on Whisper

Musixmatch's unofficial API stopped serving anonymous clients, so lyrics now come from LRCLIB — free, no account — with faster-whisper supplying the word timings; the LRCLIB words are then aligned onto those timings.

• Fixes French songs being transcribed as English: the language is now detected on voiced parts only.
• Whisper's credit hallucinations over instrumental passages are dropped.

Structure detection works again

MSAF failed to import on modern SciPy, so song sections were never detected. Sections are now labelled A, B, C… by similarity.