Steminize v0.9.2 (pre-release)

Pre-release (0.9.2) — adds an experimental native audio engine for much lower delay when you play along and record, plus a round of live-input fixes.

Steminize is a practice studio for musicians. Split any song into stems (vocals, drums, bass, guitar, piano and more) on your own computer with Demucs, then practise along: mute or solo parts, loop a section, slow it down or change the key, see the chords and notes, and play your own instrument over the tracks. Learn guitar and bass parts from tab that follows the song, with a trainer that listens to you play and a speed trainer that only speeds up when you're getting it right. Everything runs on your device; nothing is uploaded.

New: native audio engine (experimental, desktop app only)

Until now all audio went through the app's web view, which adds a lot of delay (about 240 ms round trip measured on Linux with a USB interface). The new engine talks to your sound card directly, with a buffer size you choose, so playing and recording can be tight (about 10 ms measured through PipeWire).

• Turn it on: Live input drawer, Native audio (experimental) row. Pick your output, tick Use native playback, then reopen the song. The panel marks the recommended device for your system.
• What runs natively: playback of songs saved in your library (play, loop, tempo and pitch, mixer, EQ, count-in and click, the speed trainer), your live input with monitoring, the meter, the tuner and the tab trainer, and recording. Takes line up with the song on their own: press Measure once with an output patched to an input, and takes land within about a millisecond.
• Safe to try: if the sound device is lost the app switches back to the normal engine and carries on from the same place. Turning the tick off returns everything to how it was. Songs not saved in the library still use the normal engine.
• Time-stretch sounds slightly different at slow tempos (a different algorithm).
• Not tested by the maintainer on Windows or macOS (it builds and its tests pass there, but nobody has played through it yet), and not yet on the Linux AppImage. Reports welcome.

Also in this release

• Live input: pick the input channel (Input 1 / Input 2 / both summed), which fixes a quiet signal from interfaces that leave one channel empty. The level slider goes higher. Recording, the meter, the tuner and the trainer work with Level at 0, for hearing yourself through the interface's direct monitoring.
• Lining up takes in the normal engine: a Low-latency audio option, and Line up buttons under a take to slide it against the new beat lines on every track, then Use as latency.
• Beat lines on every track once you zoom in (where the click sounds).
• Interface text can no longer be selected by dragging; no more ghost image when resizing a track.

System requirements

• Windows: 10 or 11 (.exe or .msi).
• macOS: Intel or Apple Silicon (.dmg).
• Linux: the .deb for Ubuntu 22.04 or later, Linux Mint 21 or later and Debian 12 or later (it needs libwebkit2gtk-4.1, which older releases don't have); the .rpm for Fedora-type systems; the AppImage for other distributions.
  • Songs in WAV and FLAC open out of the box. To import MP3 or AAC/M4A on Linux, the .deb/.rpm need gstreamer1.0-libav installed (sudo apt install gstreamer1.0-libav); the AppImage includes it.
  • Native audio uses ALSA. On PipeWire or PulseAudio systems it works through their ALSA plug-ins (pipewire-alsa / libasound2-plugins), which most desktops have.