Noise Cancellation C++

A high-performance audio noise cancellation engine written entirely in C++. It uses a hybrid approach, combining traditional Digital Signal Processing (DSP) techniques with a trained PyTorch Machine Learning model to intelligently remove background noise from audio files.

The system is designed for efficiency and real-time processing capabilities, making it a suitable core for applications requiring low-latency audio enhancement.

Core Features

• Hybrid Processing: Leverages a traced_denoiser_model.pt (a TorchScript model exported from Python) to guide a DSP-based spectral subtraction algorithm.
• Intelligent Adaptation: Analyzes audio in real-time to determine characteristics like Signal-to-Noise Ratio (SNR) and noise type, adapting its processing strategy accordingly.
• Frame-Based, Real-Time Capable: Processes audio in small chunks with very low latency (~1-2 ms per frame), suitable for live audio streams.
• High Performance: Built in C++ and linked against the LibTorch library for maximum computational efficiency.
• Modular Design: Separates concerns into distinct modules for audio I/O, VAD, DSP, and ML processing.

Dependencies

To build and run this project, you will need:

• A C++17 compliant compiler (e.g., Clang, GCC)
• CMake (version 3.15 or later)
• LibTorch: The C++ distribution of PyTorch. Download here.
• libsndfile: A library for reading and writing audio files.
  • On macOS: brew install libsndfile
  • On Debian/Ubuntu: sudo apt-get install libsndfile1-dev

Build Instructions

1. Download and setup LibTorch: Download LibTorch from the official PyTorch website. After extracting the archive:

   For macOS users: Remove quarantine attributes to prevent security warnings:

   xattr -r -d com.apple.quarantine "/path/to/libtorch"

2. Configure CMakeLists.txt: Open the CMakeLists.txt file and set the CMAKE_PREFIX_PATH variable to the absolute path of your unzipped LibTorch installation directory.

   set(CMAKE_PREFIX_PATH "/path/to/your/libtorch")

3. Create a build directory:

   mkdir build
   cd build

4. Generate the build system:

   cmake ..

5. Compile the project:

   make

   An executable named NoiseCancellation will be created inside the build directory.

Usage

Run the executable from the project root directory, providing an input file and an output file as arguments.

# General Usage
./build/NoiseCancellation <path/to/input.wav> <path/to/output.wav>

# Example
./build/NoiseCancellation test_audio/noisy_white.wav enhanced_audio.wav

(base) ➜  build cd ..                                                                                   
./build/NoiseCancellation test_audio/noisy_white.wav final_enhanced_output.wav
Enhanced Hybrid Noise Cancellation System
==================================================
Input file: test_audio/noisy_white.wav
Output file: final_enhanced_output.wav
Processing mode: balanced
ML model: traced_denoiser_model.pt
==================================================
Loaded audio file: test_audio/noisy_white.wav (144000 samples, 48000 Hz)
ML model loaded successfully from: traced_denoiser_model.pt
Enhanced Hybrid Noise Processor Initialized with Intelligence.
Processor initialized with 'balanced' mode.

Processing 280 frames with intelligent hybrid system...

Frame 0 analysis:
  SNR: 10.0 dB
  Noise type: broadband
  Strategy: balanced
  DSP weight: 0.30
  ML weight: 0.70
  Processing time: 22.88 ms

Frame 1 analysis:
  SNR: 0.6 dB
  Noise type: broadband
  Strategy: aggressive
  DSP weight: 0.50
  ML weight: 0.50
  Processing time: 0.53 ms

Frame 2 analysis:
  SNR: 1.1 dB
  Noise type: broadband
  Strategy: aggressive
  DSP weight: 0.50
  ML weight: 0.50
  Processing time: 0.50 ms
  - Processed frame 200/280 (71%) 

Processing completed in 848 ms.
Average processing time per frame: 3.03 ms

============================================================
INTELLIGENT PROCESSING ANALYSIS
============================================================
Performance Metrics:
  Average processing time: 1.23 ms per frame
  Average estimated SNR: 1.36 dB
  Average noise reduction: 25.87 dB
  Total frames processed: 280

Intelligent Fusion Analysis:
  Average DSP weight: 0.46
  Average ML weight: 0.54

Strategy Distribution:
  aggressive: 279 frames (99.64%)
  balanced: 1 frames (0.36%)

Noise Type Distribution:
  broadband: 267 frames (95.36%)
  mixed: 13 frames (4.64%)
============================================================

Enhanced audio saved to: final_enhanced_output.wav
Detailed metrics saved to: final_enhanced_output_metrics.json

Performance Assessment:
Real-time capable (< 10ms per frame)
Adaptive strategy selection active
Multiple noise types detected

Enhanced hybrid processing complete!
Files generated:
  - Enhanced audio: final_enhanced_output.wav
  - Processing metrics: final_enhanced_output_metrics.json

The system will process the input audio and save the enhanced version to the specified output path.

---

C++ Hybrid Noise Cancellation: Performance Comparison

The following table compares the performance of the C++ noise cancellation engine on two distinct audio files: one with difficult, broadband white noise and another with more structured, low-frequency traffic noise. The results demonstrate the system's ability to analyze the input and adapt its processing strategy accordingly.

Metric	noisy_white.wav (Difficult)	noisy_traffic.wav (Moderate)	Analysis of aaptive Behavior
Input Analysis
Average Estimated SNR	1.36 dB	4.08 dB	Correct Diagnosis: The system accurately identified the white noise as having a significantly lower Signal-to-Noise Ratio.
Dominant Noise Type	Broadband (95%)	Low-Frequency (99%)	Accurate Classification: The engine correctly profiled the noise in each file, which is crucial for selecting the right strategy.
Strategy & Performance
Strategy Distribution	99.6% Aggressive	54% Aggressive, 46% Balanced	Intelligent Adaptation: The system went "all-in" on the difficult noise but used a more nuanced, quality-preserving approach on the traffic noise.
Average Noise Reduction	25.87 dB	12.64 dB	Effective & Appropriate Action: Achieved outstanding reduction on the difficult file by being aggressive. Applied a more moderate reduction on the traffic noise to protect speech fidelity.
Average Processing Time	1.23 ms / frame	1.14 ms / frame	Excellent Performance: Both scenarios are processed nearly 10x faster than the typical real-time requirement of ~10ms.
Intelligent Fusion
Average DSP / ML Weight	0.46 / 0.54	0.52 / 0.48	Dynamic Weighting: The system subtly shifted its trust, leaning more on ML for broadband noise and slightly more on DSP for structured traffic noise.

---

I ported the core intelligence from the Python prototype to a high-performance, real-time capable C++ application.