FM / White Noise Classifier

A signal classification tool that listens through your microphone and determines whether the audio resembles an FM radio signal or white noise — using autocorrelation, FFT, and vector norm comparison.

Built with Python · NumPy · librosa · sounddevice · Tkinter · Matplotlib

---

How it works

The algorithm captures audio in real time, processes it in 2-second windows, and compares the result against pre-computed reference vectors using three DSP stages:

Microphone input
      │
      ▼
 Autocorrelation (acovf)
      │
      ▼
 Fast Fourier Transform (FFT)
      │
      ▼
 Magnitude norm vector
      │
      ▼
 L1 distance comparison → FM or White Noise

The microphone signal is normalized and compared against the average vectors computed from 100 reference .wav files (50 FM · 50 WN). The class with the smaller L1 distance wins.

---

Screenshots

🟢 Recording screen

Live spectrum while audio is captured

---

📊 Classification result

Comparative autocovariance + spectrum view

---

Project structure

FM-WN-Noise-Classifier-Algorithm/
│
├── main.py                        # Entry point
│
└── src/
    ├── data/
    │   ├── FM/                    # 50 FM reference .wav files (IS01–IS50)
    │   └── WN/                    # 50 white noise .wav files  (IS51–IS100)
    │
    ├── dataset/                   # Pre-computed reference vectors (generated by training)
    │   ├── fmVector.txt           # Average norm vector — FM
    │   ├── wnVector.txt           # Average norm vector — White Noise
    │   ├── fmAcov.txt             # Average autocovariance — FM
    │   ├── wnAcov.txt             # Average autocovariance — White Noise
    │   ├── micProcessed.txt       # Last mic norm vector
    │   └── micAcov.txt            # Last mic autocovariance
    │
    ├── interface/
    │   └── interface.py           # Tkinter GUI + Matplotlib plots
    │
    ├── loaders/
    │   └── loaderDataSet.py       # Training pipeline (loads .wav → saves vectors)
    │
    ├── models/
    │   └── classifier.py          # Classification logic (L1 distance comparison)
    │
    ├── processing/
    │   └── process.py             # DSP core: autocorrelation, FFT, norm
    │
    └── utils/
        └── audioTransformer.py    # Thread-safe mic buffer + windowed processing

---

Requirements

Package	Purpose
numpy	Vector math and FFT
librosa	Audio loading and resampling
sounddevice	Real-time microphone input
statsmodels	Autocovariance computation (acovf)
matplotlib	Signal plots embedded in GUI
tkinter	GUI framework (included with Python)

Install all dependencies:

pip install numpy librosa sounddevice statsmodels matplotlib

Note: tkinter ships with standard Python on Windows and macOS. On Linux you may need sudo apt install python3-tk.

---

Getting started

1 — Clone the repository

git clone https://github.com/your-username/FM-WN-Noise-Classifier-Algorithm.git
cd FM-WN-Noise-Classifier-Algorithm

2 — (Optional) Re-train the reference vectors

The src/dataset/ folder already contains pre-computed vectors, so this step is only needed if you want to regenerate them from the raw .wav files:

python -c "from src.loaders.loaderDataSet import train; train()"

This reads all 100 .wav files, computes the average autocorrelation + FFT norm for each class, and saves the results to src/dataset/.

3 — Run the application

python main.py

---

Usage

1. Click Empezar a grabar — the app opens your microphone and starts streaming.
2. Hold a signal source (FM radio, a phone playing static, etc.) near the mic for at least 2 seconds.
3. The live spectrum updates in real time as fragments are processed.
4. Click Detener — the app classifies the audio and shows:
   • The classification label (FM or White Noise)
   • L1 distances to each reference class
   • A two-panel comparative plot: autocovariance (top) and spectrum (bottom)

---

Signal processing details

Stage	Function	Details
Autocovariance	statsmodels.tsa.stattools.acovf	FFT-accelerated, zero-meaned
Spectrum	numpy.fft.fft applied to autocovariance	Converts time-domain pattern to frequency
Feature vector	numpy.abs(fft)	Magnitude only (phase discarded)
Comparison	Mean absolute error (L1)	Applied on normalized vectors

Audio is captured at 44 100 Hz, processed in 2-second windows (88 200 samples each). Multiple windows are averaged to reduce noise before classification.

---

License

This project is released for academic and educational use. See LICENSE if present, or contact the author for terms.