Active Noise Cancelling (Real-Time Mic Noise Suppression)

Low-latency, real-time microphone noise suppression in Python using spectral subtraction + adaptive noise tracking and 50% overlap-add (OLA). Works cross-platform with PortAudio via sounddevice.

https://github.com/SoheilGtex

---

✨ Features

• Real-time mic noise suppression (20 ms frames, 50% overlap).
• Adaptive noise estimate using exponential moving average (EMA).
• Spectral subtraction + Wiener-style gain, smoothed per band.
• High-pass filter (optional) to reduce rumble / hum.
• Calibration step (1–2 s) to capture baseline ambient noise.
• Low latency (frame = 20 ms, hop = 10 ms).
• Pure Python + NumPy/SciPy. No heavyweight ML runtime required.

This is noise suppression (post-filtering of mic input), not feedforward/feedback “anti-noise” (phase-inversion) for headphones.

---

📦 Install

git clone https://github.com/SoheilGtex/active-noise-cancelling.git
cd active-noise-cancelling
python -m venv .venv && source .venv/bin/activate  # (Linux/macOS)
# .venv\Scripts\activate                            # (Windows)
pip install -r requirements.txt

Dependencies

sounddevice (PortAudio binding)

numpy

scipy

pyyaml (for reading config.yaml, optional)

Windows: if PortAudio devices are not listed, update your audio drivers or install WASAPI loopback support (typically not needed). Linux: ensure your user is in audio group and PulseAudio/PipeWire is running. macOS: grant microphone permission to the terminal/IDE.

▶️ Run

Quick start (defaults are fine):

python main.py

Useful options

python main.py
--samplerate 16000
--frame_ms 20
--calib_sec 1.0
--device_in default
--device_out default
--highpass 80

--samplerate : 16000 or 48000 are common.

--frame_ms : 20 ms frame (10 ms hop with 50% overlap).

--calib_sec : initial ambient-noise calibration duration (seconds).

--highpass : cut below N Hz (0 to disable).

Press Ctrl+C to stop.

⚙️ How it works (DSP)

Framing & Windowing Signal is split into frames (e.g., 20 ms) with 50% overlap and Hann analysis/synthesis windows (perfect reconstruction with OLA).

Noise Spectrum Estimation (EMA) Magnitude spectrum of noise is tracked with an exponential moving average. During low-energy segments it adapts faster.

Spectral Subtraction + Gain Smoothing We estimate clean magnitude per bin: |X|_clean = max(|X| - β·|N|, floor·|N|). Then we compute a Wiener-like gain and smooth it over time to avoid musical noise.

High-Pass Optional IIR HPF removes rumble.

OLA Synthesis Inverse FFT → window → overlap-add. The output chunk each hop is sent to the soundcard with minimal latency.

📁 Config

config.yaml (optional; args override file):

samplerate: 16000 frame_ms: 20 calib_sec: 1.0 highpass_hz: 80 noise_beta: 1.0 noise_floor: 0.02 ema_alpha: 0.96 gain_smooth: 0.8 device_in: default device_out: default

🧪 Benchmark notes

Latency ≈ frame_ms (20 ms) because hop is 10 ms and OLA buffering is 1 frame.

CPU < 10% on mid-range laptops at 16 kHz mono.

📜 License

MIT

🙌 Acknowledgements

sounddevice (PortAudio) for reliable cross-platform I/O.

Classic spectral subtraction literature and Wiener filtering techniques for single-channel speech enhancement.

---