Neural Stem Slicer

Note: For development details and contribution guidelines, see Developer README

If you find this program useful please consider donating, currently looking for full time work so anything helps thanks!

Directory

• Overview
• Quick Start
• Low Level Architecture
• Installation
  • Prerequisites
  • macOS Installation
  • Windows Installation
• Usage
• Updating

Overview

A music production tool / powerful folder that converts any song into 8 stems (bass, vocal, other(melodic), drums, kick, snare, tom, cymbal) labeled by BPM and Key, and then each stem into into 8 bar segments.

          ┌─────────────────-─┐       ┌─────────────────────┐      ┌─────────────────────┐
          │ Module 1          │       │  Module 2           │      │   Module 3          │
 Song ──> │ BPM Key Analysis  │  ──>  │  8 Stem Separation  │ ──>  │  8-Bar Segments     │  ──> Segments/Stems  
          │ (DeepRhythm)      │       │  (2 stage Demucs)   │      │  (SoundFile)        │
          │ (Librosa)         │       │                     │      │                     │
          └─────────────────-─┘       └─────────────────────┘      └─────────────────────┘

Quick Start

1. After installation, place any .mp3 or .wav file in the root folder
2. Double click split_stems.command to run the process
3. Neural Stem Slicer will automatically:
   • Split tracks into 8 high-quality stems using Demucs Hybrid Transformers
     • Bass, Vocals, Melody, Drums
     • Drums → Kick, Snare, Toms, Cymbals
   • Detect and label all files with BPM and key using deep learning analysis (with manual override)
   • Segment each stem into precise 8-bar loops ready for Ableton Complex warping
4. All output files should be in the /output folder

Perfect for:

Music Production & Sampling/Hardware

• Create custom loops for Ableton Session view
• Extract clean stems for remixing and sampling
• Build your own custom loop collections labeled by BPM/Key
• Extract usable stems from vinyl rips
• Load stems into hardware (samplers, drum machines, groove boxes)
• Study arrangements by analyzing each 8 bar segment

Live Performance & DJing

• Generate clean acapellas and instrumentals
• Create perfectly-timed drum loops for live layering
• Build custom DJ tools and transition elements
• Study arrangement and mixing techniques
• Extract vocals and music beds for content creation

Low Level Architecture

---

Installation

Prerequisites

• Git (Download Git)
  • macOS: brew install git
  • Windows: Download installer from git-scm.com
• Python 3.11 (Download Python)
  • macOS: brew install python@3.11
  • Windows: Download installer from python.org
• macOS (including Apple Silicon), Linux, or Windows

Setup

# Clone repository
git clone https://github.com/dennisrcao/neural-stem-slicer.git
cd neural-stem-slicer

macOS Installation

# Verify Python installation
python3.11 --version  # Should show Python 3.11.x

# Create virtual environment
python3.11 -m venv .venv

# Activate virtual environment
source .venv/bin/activate

# Install dependencies
pip install -r requirements.txt

# Install tkinter for Python 3.11
brew install python-tk@3.11

# Install drumsep model
python step3_0_Seperation_Models/drumsep/drumsepInstall.py

# Make it executable
chmod +x step3_0_Seperation_Models/drumsep/drumsep

Windows Installation

# Install Python 3.11 from python.org

# Verify installation
python --version  # Should show Python 3.11.x

# Create virtual environment
python -m venv .venv

# Activate virtual environment
.venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

# Install drumsep model
python step3_0_Seperation_Models/drumsep/drumsepInstall.py

---

Usage

After you've finished the setup installation process, you can either double click the file split_stems.command or in your terminal run

python split_stems.py

This should bring up the GUI which I may get rid tbh.

For now you can select which steps/modules it goes through with the checkboxes.

Module 1: I calculate the BPM and Key, if they are incorrect you can enter in a value into the Manual Override box

Module 2: Stem Seperation - 8 stage demucs stem seperation.

Module 3: Segmentation - Chops each 8 step into perfect 8 bar segments (based on knowing the correct BPM). ❗Currently you need to export your song starting right at the "1" beat as I'm still in the process of implementing on beat detection

---

Updating

I will occasionally be pushing updates to this repository, to make sure you have the latest version follow the steps:

# cd into the project repo
cd neural-stem-slicer
# pull down latest changes
git pull origin main