Dredge Looper

A practice looper for Linux: load a song, loop a section, slow it down without changing pitch, and drill it until you can play it.

Releases · Features · Dependencies · Install · Build & develop

---

Features

Basic

These work with the installed app — no ML setup.

Playback

• Sample-accurate looping — set a loop by dragging on the waveform; the seam is crossfaded. Loops are saved per song and take their names from the sections they span (verse 2 → chorus 1).
• Pitch-preserving speed — 0.25–2.0× via Rubber Band R3 (compiled in); independent pitch shift, ±12 semitones plus cents.
• Bass focus — octave-up plus low-pass to isolate basslines.

Practice

• Drill — tempo trainer that raises speed across passes, with region shaping and a recall mode that mutes playback so you play from memory.
• Practice routines — saved sequences of practice blocks; each block sets a loop span, mix, speed, lead-in, and count-in, and the app steps through them as you play.
• Overdub recording — record yourself (mic or audio interface) over the song, the selection, or the active loop. Recordings are latency-calibrated, stored in the song bundle, and play back over the mix.
• Metronome and count-in — manual BPM, or synced to the analyzed BPM once a song is analyzed.
• Tuner — chromatic tuner with note and cents and a hold-to-lock confirm. Works with no song loaded.

Songs and notes

• Sections and notes — mark sections by hand; per-section free text with inline tablature, keyed to the section occurrence (verse 2).
• Markers — set positions in the song and jump playback to them.
• Song bundles — each song is a self-contained directory (audio + dredge.json holding sections, loops, notes, analysis, recordings). Copy the folder to another machine and it loads with everything.
• Wide format support — imports mp3, flac, ogg, opus, wav, and m4a, and takes the audio track from mp4/mov/webm/mkv video files (opus, webm, and mkv via ffmpeg).
• Export — render the current mix (stem balance, speed, pitch, bass focus) to WAV, or MP3 with ffmpeg.

Control and interface

• MIDI foot pedal — map pedals to transport, markers, and isolation snapshots with a learn flow, hands on the instrument.
• Dock layout — the right-hand tabs (structure, loops, routines, export, …) live in resizable panels; drag tabs to reorder, merge, or split panels.
• Control socket — every command the UI uses is also available as JSON over a Unix socket; a headless daemon (dredged) runs the same engine without the UI.

With ML enabled

These require the optional Python tools in Dependencies.

• Detected song structure — beats, downbeats, BPM, and labelled sections detected and drawn on the waveform.
• Downbeat snapping — loop and selection edges snap to detected downbeats.
• Section click track — a click on the analyzed beats inside sections you choose, accented on downbeats.
• Stems — 6-stem separation (vocals / drums / bass / guitar / piano / other) with per-stem faders. Runs locally.

Dependencies

Basic

The stretch engine (Rubber Band) is compiled into dredge, so there is no rubberband package to install. The prebuilt binaries target Debian/Ubuntu library versions — on Arch, use the dredge AUR package.

ML enabled

All ML pieces require uv on PATH: sudo pacman -S uv, or on Ubuntu curl -LsSf https://astral.sh/uv/install.sh | sh. A GPU is optional throughout — CPU works, slower.

Beat / section analysis (dredge-enable-ml analyze)

• venv: ~/.local/share/dredge/analyze-venv, Python 3.12 (override path with $DREDGE_ANALYZE_VENV)
• packages: beat_this (from git), torch, soundfile, librosa, einops, rotary-embedding-torch
• provides: beat / downbeat / BPM grid (beat_this) and novelty-based section boundaries
• disk: torch download, several GB

Higher-quality sections (dredge-enable-ml songformer)

• venv: ~/.local/share/dredge/songformer-venv, Python 3.11 (override with $DREDGE_SONGFORMER_VENV)
• packages: torch==2.4.0, torchaudio==2.4.0, numpy<2, transformers==4.51.1, librosa, soundfile, ema-pytorch, loguru, omegaconf, tqdm, safetensors, muq, x-transformers, msaf, einops, huggingface_hub
• also downloads the ASLP-lab/SongFormer model snapshot from Hugging Face on first run (weights plus its own modeling code)
• runs alongside the beat grid, so it also needs the analyze venv above
• VRAM at run time: ~8 GB resident, brief peak up to ~15 GB. Falls back to the novelty detector if the venv is absent or the run runs out of memory.

Stem separation (dredge-enable-ml stems)

• installed as a uv tool: uv tool install demucs --with torchcodec
• model: htdemucs_6s, the 6-source Hybrid Transformer Demucs (Meta AI); weights download on first run
• provides: 6-stem separation (vocals / drums / bass / guitar / piano / other)
• needs ffmpeg (above) for stem export
• disk: PyTorch, ~2.5 GB

Install

Linux only. The audio engine is PipeWire-native: PipeWire 1.0+ is required, with no ALSA or PulseAudio fallback.

Basic

Arch / Arch-based

yay -S dredge   # builds from source against your system libraries

Debian / Ubuntu (24.04+ / Debian 13+)

Download the latest dredge_*_amd64.deb from the releases page, then:

sudo apt install ./dredge_*_amd64.deb

apt pulls the runtime libraries automatically. The basic features above run with nothing else installed.

ML enabled

Beat/section analysis and stem separation are off by default and self-bootstrap on first use. dredge-enable-ml does that bootstrap up front, so the multi-GB downloads happen now instead of on the first analysis:

dredge-enable-ml all          # analyze + songformer + stems
dredge-enable-ml analyze      # beat/section analysis only
dredge-enable-ml songformer   # higher-quality section labels
dredge-enable-ml stems        # stem separation only

Checking a setup

dredge-doctor reports which optional tools are installed and the exact command to add each missing one. The desktop app shows the same under Settings → capabilities.

Status

Dredge has only been used on two machines, both running Arch Linux. Assumptions that hold there may break elsewhere — report an issue for anything you hit.

Development

Built with Rust, Tauri 2, and Svelte 5. Building from source and hacking on it are covered in DEVELOPMENT.md.

MIT licensed; the binaries bundle the Rubber Band Library (GPL-2.0-or-later), so distributed builds are GPL-governed.