Control individual elements — vocals, drums, bass, instruments — independently.
Built-in AI stem separation, ML-powered analysis, and smart track suggestions.
Mesh is a DJ software suite for mixing electronic music by stems. Import any audio file and Mesh separates it into four stems (vocals, drums, bass, other) using the Demucs neural network. Then mix, apply effects, and perform — with independent control over each element.
Automatic beat sync and loudness matching make it approachable for beginners, while deep per-stem effect routing and plugin hosting satisfy experienced DJs looking to go beyond what commercial software offers.
| Application | Purpose |
|---|---|
| mesh-cue | Prepare your library — import, separate stems, analyze BPM/key, edit beat grids, manage playlists, export to USB |
| mesh-player | Perform live — 4-deck stem mixing, per-stem effects, beat sync, MIDI controllers, smart suggestions, set recording |
Mesh is designed for electronic music with regular beat grids — techno, house, DnB, trance, and similar genres. The automatic beat sync and grid analysis assume consistent tempo. Genres with irregular timing (live recordings, jazz, hip-hop with swing) will need more manual grid work.
Most DJ software treats stems as an afterthought — mute or solo, maybe a basic filter. Mesh builds the entire workflow around stems, from preparation to performance.
Per-stem effect chains — Route each stem through its own multiband effect chain with up to 8 frequency bands and 8 effects per band. Reverb on vocals, compression on drums, filter sweeps on bass — simultaneously. 4 macro knobs per deck map parameters across stems for one-knob control of complex transformations. Rekordbox, Serato, and Traktor offer basic stem filtering; mesh gives you full effect routing.
CLAP plugin hosting — Load professional studio plugins (LSP, Dragonfly Reverb, Airwindows) as stem effects. No other DJ software hosts CLAP plugins.
Pure Data custom effects — Write visual audio patches that run as live stem effects with up to 8 mapped parameters. Includes support for RAVE neural audio synthesis via the nn~ external.
ML-powered library — During import, neural networks classify genre (400 Discogs categories), detect vocal presence (96% accuracy), and extract audio fingerprints for similarity search. Beat detection uses the Beat This! model (ISMIR 2024) — eliminates half-tempo errors common with DnB and fast tempos.
Smart track suggestions — Recommendations based on audio fingerprint similarity, harmonic compatibility, BPM proximity, and perceptual arousal. An energy direction fader steers suggestions toward higher or lower energy. Two key matching algorithms: classic Camelot wheel or Krumhansl perceptual model. All processing is local — no cloud, no subscription.
Stem linking — Prepare mashups by linking stems across tracks. Vocals from track A over instrumentals of track B, pre-aligned and time-stretched. Links persist in your library for instant recall during performance.
$112 embedded standalone — Run the full player on an Orange Pi 5 Pro with an I2S DAC. Boot directly into a kiosk DJ interface — no laptop, no subscription, no vendor lock-in.
Feature comparison with other DJ software
| Feature | Mesh | Rekordbox | Serato | Traktor | Mixxx |
|---|---|---|---|---|---|
| Stem separation | Built-in (Demucs) | Built-in | Built-in | Built-in (iZotope) | In development |
| Per-stem effect chains | Yes | No | No | No | No |
| CLAP plugin hosting | Yes | No | No | No | No |
| Custom PD effects | Yes | No | No | No | No |
| Neural audio effects (RAVE) | Yes | No | No | No | No |
| ML genre/mood tagging | Yes | No | No | No | No |
| Audio fingerprint suggestions | Yes | No | No | No | No |
| Perceptual key matching | Yes | No | No | No | No |
| Stem slicer | Yes | No | No | No | No |
| Embedded standalone | Yes (~$112) | No | No | No | Community |
| Vertical waveform layout | Yes | No | No | No | No |
| Streaming services | No | Yes | Yes | Yes | No |
| DVS / vinyl control | No | Yes | Yes | Yes | Yes |
| Sampler / remix decks | No | Yes | Yes | Yes | Yes |
| Video mixing | No | Add-on | Add-on | No | No |
| Auto DJ | No | No | No | No | Yes |
| macOS | Planned | Yes | Yes | Yes | Yes |
| Price | Free (AGPL) | Subscription | Subscription | One-time | Free (GPL) |
Mesh prioritizes depth of stem control and audio intelligence over breadth of DJ formats. If you need vinyl control, streaming services, or video mixing, Rekordbox/Serato/Traktor are better choices. If you want to push what's possible with stems, effects, and AI-assisted mixing — mesh is built for that.
| Package | Platform | Description |
|---|---|---|
mesh-cue_amd64.deb |
Linux (Ubuntu 22.04+) | Full suite with stem separation (CPU) |
mesh-cue-cuda_amd64.deb |
Linux (Ubuntu 22.04+) | Full suite with NVIDIA CUDA GPU acceleration |
mesh-cue_win.zip |
Windows 10/11 | Full suite with DirectML GPU acceleration |
mesh-player_amd64.deb |
Linux (Ubuntu 22.04+) | Lightweight performance player |
mesh-player_win.zip |
Windows 10/11 | Lightweight performance player |
Download from GitHub Releases.
# Full suite (CPU stem separation)
sudo dpkg -i mesh-cue_amd64.deb
# Or with NVIDIA GPU acceleration
sudo dpkg -i mesh-cue-cuda_amd64.deb
# Performance player only
sudo dpkg -i mesh-player_amd64.debRequirements: PipeWire or JACK audio server. For CUDA: NVIDIA driver 525+ with CUDA 12.
Extract the ZIP and run mesh-cue.exe or mesh-player.exe. A DirectX 12 capable GPU is recommended for faster stem separation but not required — it falls back to CPU automatically.
Launch mesh-cue and place audio files (MP3, FLAC, WAV, OGG, M4A) in ~/Music/mesh-collection/import/.
Pre-separated stems: Name files as Artist - Track_(Vocals).wav, _(Drums).wav, _(Bass).wav, _(Other).wav and select Stems import mode.
Regular audio files: Select Mixed Audio mode. Mesh separates each file into 4 stems automatically (2-5 min/track on CPU, 15-30 sec with GPU).
During import, mesh also analyzes BPM, musical key, loudness (LUFS), audio fingerprints, and optionally classifies genre and mood using neural networks.
Load a track in mesh-cue to:
- Verify and adjust BPM detection
- Fine-tune beat grid alignment (nudge or align to playhead)
- Set up to 8 color-coded hot cues at key moments
- Save loops for performance recall
- Optionally link stems from other tracks for mashups
Click Export to sync playlists to a USB drive. Select which playlists to include, and mesh copies the stems, metadata, and presets. The USB works with mesh-player on any machine or embedded device.
Launch mesh-player, browse your collection, and load tracks onto the 4 decks.
| Control | Function |
|---|---|
| Play / Cue | CDJ-style transport with cue point preview |
| Stem buttons | Mute/solo individual stems per deck |
| Beat sync | Automatic beat grid alignment to global BPM |
| Key match | Automatic harmonic pitch adjustment |
| Loop | Quantized loops from 1/8 beat to 256 beats |
| Hot cues | 8 instant-jump points per track |
| Slicer | Real-time stem slice resequencing |
| Suggestions | AI track recommendations with energy direction control |
Track Preparation (mesh-cue)
- AI stem separation — Import regular audio files and automatically separate into Vocals, Drums, Bass, and Other using Demucs. Standard or fine-tuned models, 1-5 processing shifts for quality vs speed. GPU acceleration via CUDA (Linux) or DirectML (Windows)
- BPM detection — Two backends: Simple (Essentia RhythmExtractor) for speed, or Advanced (Beat This! ONNX model) for accuracy. Advanced mode includes downbeat detection and eliminates half-tempo errors on DnB and fast tempos
- Key detection — Automatic musical key identification via Essentia
- Beat grid editing — Nudge grid in ~2.5ms increments, align to playhead position, adjust BPM. Grid updates instantly on the waveform
- 8 hot cues per track — Color-coded, beat-snapped, CDJ-style jump points
- 8 saved loops per track — Store and recall loop regions with variable length
- Stem linking — Link a stem from another track for prepared mashups with automatic time-alignment and BPM stretching. Drop markers for alignment reference
- ML audio analysis — Genre classification (400 Discogs categories), vocal/instrumental detection (96% accuracy), mood and arousal estimation, audio characteristics (timbre, tonality, danceability). All on-device using EffNet neural networks
- Batch re-analysis — Re-analyze metadata, BPM, or ML tags for selected tracks, playlists, or the entire collection
- Playlist management — Create, organize, and nest playlists with drag-and-drop between dual browser panels
- USB export — Sync playlists to USB drives with stems, metadata, and presets. Per-track progress, ETA, and cancellation support
Live Performance (mesh-player)
- 4-deck architecture — Load and mix up to 4 tracks simultaneously, each with 4 independent stems
- Per-stem control — Mute, solo, and adjust volume for each stem independently
- Automatic beat sync — Tracks phase-lock to the global BPM on play. Configurable phase sync on/off
- Automatic key matching — Pitch-shift tracks to match harmonically
- Stem slicer — 8 slice pads per stem for real-time pattern resequencing with 8 storable presets per track
- Quantized loops — 1/8 beat to 256 beats with halve/double from encoder. Beat jump forward/backward by loop length
- 8 hot cues — Instant jump points, color-coded, beat-snapped
- Waveform layouts — Horizontal (default) or Vertical (time flows top-to-bottom). Three abstraction levels (Low / Medium / High) control visual detail
- Set recording — Record master output to WAV on all connected USB sticks simultaneously. Automatic tracklist TXT file generated from session history with timestamps
- Auto-gain — LUFS-based loudness normalization with configurable target: -6 (loud), -9 (medium), -14 (streaming), -16 (broadcast)
- Collection browser — Search, sort, and browse playlists. Load tracks to specific decks. USB hot-plug detection. Played tracks dimmed
- 3-band EQ + filter — Per-deck low/mid/high EQ and sweepable DJ filter (HP/LP combo, 60 Hz - 20 kHz)
- Mixer — Per-channel volume, cue enable, master volume, cue volume, cue/master mix
Effects System
- Multiband effect container — Split each stem into up to 8 frequency bands with Linkwitz-Riley crossovers. Each band has its own effect chain (up to 8 effects per band)
- 4 macro knobs per deck — Map parameters across multiple stems and effects for one-knob control of complex transformations
- CLAP plugin hosting — Load any CLAP audio plugin as a stem effect. Plugin discovery from
~/.clap/and/usr/lib/clap/. Tested with LSP, Dragonfly Reverb, Airwindows, BYOD, ChowTapeModel - Pure Data effects — Custom PD patches with up to 8 mapped parameters per effect. RAVE neural audio synthesis supported via nn~ external
- Built-in effects — Sweepable DJ filter (HP/LP), delay (with feedback), reverb, gain
- Automatic latency compensation — Delay-line alignment across stems, bands, and effect chains. Dynamic re-compensation when plugins change latency
- Deck presets — Save and recall complete effect configurations across all 4 stems. Reference stem presets by name for reuse
See docs/effects.md for details on creating PD patches, CLAP plugin setup, and preset management.
Smart Suggestions
Toggle suggestions in the collection browser to get track recommendations based on what's currently playing.
Scoring factors:
- Audio fingerprint similarity (HNSW vector search on 16-dimensional audio features)
- Harmonic key compatibility (Camelot wheel or Krumhansl perceptual model)
- BPM proximity
- Energy direction alignment (arousal, danceability, genre-normalized aggression)
Energy direction fader — Steer suggestions toward higher-energy or cooler tracks. At center, audio similarity dominates for safe harmonic matches. At extremes, energy-aware scoring takes over with relaxed harmonic filtering, rewarding bolder key transitions.
Key scoring models (selectable in Settings):
- Camelot — Classic DJ wheel with hand-tuned transition scores
- Krumhansl — 24x24 perceptual key distance matrix from music psychology research. More nuanced with cross-mode transitions (e.g., C major to C minor)
Reason tags — Each suggestion shows a colored pill (green/amber/red) with a directional arrow indicating the harmonic relationship and match quality.
See docs/smart-suggestions.md for the full scoring pipeline, all transition types, and parameter reference.
Audio Quality
- Zero-dropout loading — Load new tracks while playing without audio glitches. Lock-free architecture with cyclic buffer pool
- Professional time stretching — signalsmith-stretch for tempo changes without pitch artifacts
- Master bus protection — Soft limiter + hard clipper prevent distortion and protect speakers
- Low-latency audio — JACK on Linux (with PipeWire bridge), WASAPI on Windows. Typical output latency: 5-12ms
- Separate master and cue outputs — Route headphones to a different audio interface or channel pair for monitoring
- Latency-compensated beat sync — Real output pipeline latency is measured from hardware timestamps and used to compensate beat-snap decisions
- Direct MIDI path — Timing-critical commands (play, cue, hot cue, beat jump) bypass the UI tick loop and go directly to the audio engine via a lock-free ring buffer
Detailed guides for each area of mesh:
| Guide | Covers |
|---|---|
| Collection & Import | Folder structure, import workflows, USB export/sync, database, set recordings |
| MIDI Controllers | Learn wizard, deck layers, LED feedback, shift modes, compact 4-deck mapping |
| Effects System | Multiband container, CLAP plugins, Pure Data patches, RAVE neural effects, presets |
| Smart Suggestions | Scoring pipeline, energy fader, key matching, transition types, reason tags |
| Configuration | All settings (player & cue), themes, audio device setup, CLI arguments |
| Embedded Standalone | Orange Pi hardware, DAC wiring, SD image, WiFi, OTA updates, troubleshooting |
Mesh works with any class-compliant MIDI controller. The MIDI Learn wizard walks you through mapping all controls — transport, stems, mixer, browser, effects, and performance pads.
| Device | Protocol | Features |
|---|---|---|
| Allen & Heath Xone K2/K3 | MIDI | Rotary encoders, buttons, note-offset RGB LEDs with beat-synced pulsing |
| Native Instruments Kontrol F1 | HID | 4x4 RGB pad grid, encoders, faders, full-color LED feedback |
| Pioneer DDJ-SB2 | MIDI | Profile included |
| Any MIDI controller | MIDI | Via Learn wizard |
Layer toggle — Map 4 virtual decks to 2 physical controller sides with a single layer switch button.
Compact mode — Momentary mode overlays let you hold a button to temporarily switch pads to Hot Cue or Slicer mode, then release to return to default stem controls.
See docs/midi-controllers.md for the learn wizard guide, LED feedback details, and configuration reference.
Run mesh-player on an Orange Pi 5 Pro (~$80) as a standalone DJ unit — no laptop required.
| Component | Role | Approx. Cost |
|---|---|---|
| Orange Pi 5 Pro 8GB | RK3588S SoC, WiFi 5, USB 3.0 | ~$80 |
| PCM5102A I2S DAC board | Master audio output (112 dB SNR) | ~$5 |
| ES8388 onboard codec | Headphone cue output (3.5mm) | Included |
| microSD card (32GB+) | NixOS system image | ~$8 |
| Enclosure + cables | Optional | ~$19 |
| Total | ~$112 |
The device boots directly into mesh-player in fullscreen kiosk mode (NixOS + cage Wayland compositor). Load tracks from USB 3.0 sticks. Over-the-air updates via WiFi.
Quick start:
- Download the SD image from GitHub Releases (look for
sdimage-*tags) - Flash:
zstdcat nixos-sd-image-*.img.zst | sudo dd of=/dev/sdX bs=4M status=progress - Boot — mesh-player starts automatically
See docs/embedded.md for the full hardware guide with wiring diagrams, device tree configuration, audio routing, WiFi/network setup, and OTA deployment.
Mesh stores its collection and configuration in ~/Music/mesh-collection/:
mesh-collection/
├── import/ # Drop files here for import
├── tracks/ # Stem library (8-channel FLAC)
├── playlists/ # Playlist folders
├── presets/ # Effect presets (stems/ and decks/)
├── effects/ # Pure Data effect patches
├── config.yaml # Application settings
└── theme.yaml # Color theme
Themes — 5 built-in color themes (Mesh, Catppuccin, Rose Pine, Synthwave, Gruvbox) with customizable stem colors and UI accents. Themes switch instantly without restart.
Fonts — Exo (default), Hack, JetBrains Mono, Press Start 2P. Font and size changes require restart.
See docs/collection.md for details on import workflows, USB export, database, and set recordings. See docs/configuration.md for the complete settings reference, theme customization, and audio device setup.
Features in development or planned:
- macOS support
- Tag editing UI (add/edit/remove tags from the browser)
- Auto headphones cue (volume-based automatic cue routing)
- Live peak meters (per-channel and master)
- Built-in native effects (beat-synced echo, flanger, phaser, gater)
- Database versioning (schema migration for USB forward-compatibility)
- Session history browser and set reconstruction
- Slicer morph knob (scroll through preset banks)
- Jog wheel beat nudging
- B2B networked mode (dual-system Ethernet link)
git clone https://github.com/dataO1/Mesh.git
cd Mesh
nix develop # Enter dev environment with all dependencies
cargo build --release -p mesh-player
cargo build --release -p mesh-cueRequirements: Nix package manager — handles all native dependencies (Essentia, FFmpeg, JACK, Vulkan, hidapi, etc.).
See ARCHITECTURE.md for technical documentation on the audio engine, real-time architecture, and crate structure.
Mesh is not accepting code contributions at this time. If you find a bug or have a feature request, please open an issue.
AGPL-3.0 — see LICENSE.
Uses Essentia (AGPL-3.0) for audio analysis and Demucs (MIT) for stem separation. Genre and mood classification models from the Essentia model hub (CC BY-NC-SA 4.0). Beat detection model from Beat This! (MIT).
- Demucs — Neural stem separation (Meta AI)
- signalsmith-stretch — Time stretching
- Essentia — Audio analysis (key, LUFS, BPM)
- Beat This! — ML beat tracking (CPJKU, ISMIR 2024)
- iced — GUI framework
- JACK — Professional audio routing
- CozoDB — Embedded graph database
- ort — ONNX Runtime bindings for Rust
- Pure Data — Visual audio programming
- RAVE — Real-time Audio Variational autoEncoder