Spectral Subtractor

A Real-Time Spectral Subtraction VST3 Plugin & Audio Analysis Study.

🎓 About The Project

Spectral Subtractor was developed as a deep-dive educational project to master Digital Signal Processing (DSP), C++ Audio Architecture, and the JUCE Framework.

While the primary goal was to create a functional, low-latency noise reduction tool for live streaming (OBS Studio), the core motivation was to demystify how audio analysis works "under the hood." It implements a custom spectral subtraction engine from scratch—capturing environmental noise profiles, visualizing them via a scrolling spectrogram (waterfall plot), and mathematically removing them from the signal chain without destroying voice fidelity.

🛠️ Technical Stack

• Language: C++17
• Framework: JUCE 7
• DSP Concepts: Fast Fourier Transform (FFT), Windowing, Overlap-Add (OLA), Ring Buffers.
• Architecture: Thread-safe Consumer/Producer pattern (Audio Thread vs. GUI Thread).

🚀 Key Features

• Adaptive Noise Profiling: "Learn" mode uses averaging logic to capture a stable "fingerprint" of room tone (fans, AC) while ignoring transient spikes.
• Real-Time Visualization: High-performance scrolling heatmap allows users to see the frequency content of their audio history.
• Visual Analysis Tools:
  • Variable Zoom: Focus inspection on specific bands (e.g., Sub-bass vs. Human Speech).
  • Visual Gain: Post-process the visual data to reveal low-amplitude noise floors without altering the audio output.
  • A/B Testing: Bypass toggles to audit the subtraction algorithm in real-time.
• Broadcast Safe: Optimized for zero-latency monitoring with CPU-efficient circular buffering.

🧠 What I Learned (Engineering Highlights)

This project served as a practical application of several advanced audio programming concepts:

1. Digital Signal Processing (DSP)

• The Math of Sound: Moved beyond time-domain amplitude to frequency-domain analysis using FFT.
• Windowing: Implemented Hanning windows to mitigate spectral leakage and "clicking" at buffer boundaries.
• Phase Coherence: Learned the critical importance of preserving phase data during reconstruction to avoid robotic/metallic artifacts (the "underwater" effect).

2. Low-Level Memory Management

• Ring Buffers: Built a custom circular buffer implementation to bridge the gap between the host's audio block size and the FFT's required window size.
• Crash Safety: Solved critical "Off-by-One" memory access violations common in signal processing loops, ensuring stability even during host initialization.

3. VST Architecture

• Thread Safety: Strict separation of the Processor (High-priority audio thread) and the Editor (Low-priority UI thread).
• Plugin Lifecycle: Managing state, parameters, and entry points (createPluginFilter) for VST3 host compatibility.

🛠️ Development & Build Instructions

This repository tracks the source code and Projucer project file. JUCE Library Code and Build artifacts are ignored to keep the repository clean.

To build the project:

1. Prerequisites: Install JUCE 7 and ensure you have the Projucer application.
2. Generate Project Files:
   • Open SpectralSubtrator.jucer in the Projucer.
   • Click File > Save Project (or Ctrl + S) to regenerate the JuceLibraryCode and Builds folders for your platform (VS2022, Xcode, etc.).
3. Compile:
   • Open the generated solution in your IDE.
   • Build the VST3 target in Release mode.

📥 Installation

1. Download the latest release.
2. Copy SpectralSubtractor.vst3 to your system's VST3 folder:
   • Windows: C:\Program Files\Common Files\VST3\
3. Load in OBS Studio:
   • Note: This plugin requires a VST3 wrapper to function correctly in OBS. I recommend using atkAudio or Element.
   • Right-click your Audio Source in the Mixer -> Filters.
   • Click + and add your VST Wrapper (e.g., Vst 2.x Plug-in -> select atkAudio).
   • Open the wrapper interface and load SpectralSubtractor from the plugin list.

🔮 Future Improvements

• Transient Shaping: Implementing transient detection to bypass noise reduction during sharp sounds (drum hits, consonants) to preserve attack.
• Pitch Shifting: Utilizing the existing frequency bins to implement pitch correction or alteration.
• Neural Network Integration: Replacing the static subtraction threshold with a lightweight ML model for dynamic voice activity detection (VAD).

👤 Author & Contact

Alec Brandt (Azterisk) - Student & Software Developer

I am currently a student focusing on Software Testing, Systems Analysis, and C++ Development. I am actively looking for opportunities to apply my skills in audio software, tool development, or systems engineering.

---

Built with JUCE. MIT License.