AudioForge

Low-latency microphone audio processor with AI noise suppression and 10-band parametric EQ.

Forge your sound in real-time. Built for streamers, content creators, and anyone who wants professional microphone processing without the complexity.

Features

• AI Noise Suppression - Three models to choose from:

  • RNNoise (~10ms latency)
  • DeepFilterNet LL (~10ms latency, better quality)
  • DeepFilterNet Standard (~40ms latency, best quality)

• 10-Band Parametric EQ - Shaping with presets:

  • Voice, Bass Cut, Presence, Warm & Clear
  • Per-band frequency, gain, and Q control

• Full DSP Chain - All in one place:

  • Noise Gate with IIR envelope follower
  • AI Noise Suppression (selectable model)
  • 10-Band Parametric EQ
  • Compressor with soft-knee and makeup gain
  • Hard Limiter for clipping prevention

• Real-Time Monitoring - See what's happening:

  • OBS-style visual level meters (input/output)
  • DSP performance metrics (latency, processing time)
  • Buffer health monitoring
  • Compressor gain reduction meter

Screenshots

Coming soon

Quick Start

Prerequisites

• Python 3.9+
• Rust 1.70+ (with maturin)
• Audio interface (VB Audio Cable recommended for routing)

Installation

1. Clone the repo:

git clone https://github.com/FueledByRedBull/audio-forge.git
cd audio-forge

2. Build the Rust core:

By default, ALL features are enabled (RNNoise + VAD + DeepFilterNet):

cd rust-core
maturin develop --release

To build with only specific features, edit pyproject.toml:

[tool.maturin]
features = ["pyo3/extension-module"]  # Add "vad", "deepfilter" as needed

To include this...	Add to features line
VAD (smart gate)	vad
DeepFilterNet	deepfilter
Both	vad, deepfilter

Examples:

# VAD only
features = ["pyo3/extension-module", "vad"]

# DeepFilterNet only
features = ["pyo3/extension-module", "deepfilter"]

# Both (same as default)
features = ["pyo3/extension-module", "vad", "deepfilter"]

# Minimal (RNNoise only)
features = ["pyo3/extension-module"]

3. Run the application:

python -m mic_eq

---

Optional: DeepFilterNet Support

DeepFilterNet requires both the C library and model file.

Step 1: Build the C library (df.dll)

# Clone DeepFilterNet
git clone https://github.com/Rikorose/DeepFilterNet.git
cd DeepFilterNet

# Build the C library
cargo build --release

# Copy the DLL to AudioForge
cp target/release/df.dll ../audio-forge/

Step 2: Download the model

# Create models directory in AudioForge
mkdir audio-forge/models

# Download DeepFilterNet3 LL model (recommended for real-time)
# Place in: audio-forge/models/DeepFilterNet3_ll_onnx.tar.gz

Or set DEEPFILTER_MODEL_PATH environment variable to point to your model file.

---

Optional: Silero VAD (for VAD-assisted gate)

VAD-assisted gate modes require the Silero VAD model:

# Create models directory in AudioForge
mkdir audio-forge/models

# Download Silero VAD model
# Visit: https://github.com/snakers4/silero-vad/tree/master/files
# Download: silero_vad.onnx (v4 or v5)
# Place in: audio-forge/models/silero_vad.onnx

Or set VAD_MODEL_PATH environment variable.

Usage

1. Select Audio Devices

   • Input: Your microphone
   • Output: VB Audio Cable (for routing to Discord/OBS)

2. Choose AI Model

   • RNNoise - Fastest, good quality
   • DeepFilterNet LL - Best for real-time
   • DeepFilterNet - Best quality, higher latency

3. Adjust Settings

   • Use presets or tweak individual bands
   • Adjust gate threshold, compressor, limiter as needed

4. Start Processing

   • Click "Start Processing"
   • Monitor levels in real-time

Roadmap

v1.5.0 (Current Release)

• RNNoise integration
• Silero VAD-assisted noise gate
• DeepFilterNet integration (experimental, C FFI)
• 10-band parametric EQ
• Noise gate, compressor, limiter
• Real-time level meters
• DSP performance metrics
• Preset system
• Scrollable panels
• Adaptive release compressor

v1.6.0 (Planned)

• Linux/macOS builds
• Installation script
• More EQ presets
• VST plugin format

v2.0.0 (Future)

• WebRTC/NDI support
• Multiple input/output profiles
• Scene-based preset switching
• Cloud preset sharing

Architecture

Processing Chain:

Mic Input → Pre-Filter (DC Block + 80Hz HP) → Noise Gate → AI Noise (RNNoise/DeepFilter) → 10-Band EQ → Compressor → Limiter → Output

Note: DeepFilterNet is experimental. RNNoise is recommended for production use.

Target Latency: <30ms total (DeepFilterNet LL), <50ms (Standard)

License

MIT License - see LICENSE file.

Acknowledgments

• Inspired by SteelSeries GG Sonar's ClearCast AI
• RNNoise by Jean-Marc Valin
• DeepFilterNet by Hendrik Schröter
• Spectral Workbench (reference implementation)

---

Made with ❤️ by FueledByRedBull