Noise Canceller

How the Noise Canceller Works

The Noise Canceller uses your browser's Web Audio API with AudioWorklet processors running on a dedicated audio rendering thread. It captures ambient sound through your microphone, analyses it in real time, and generates both adaptive masking noise and active anti-noise through your headphones.

Masking Noise Types

• White noise — flat spectrum, masks high-pitched and variable sounds.
• Pink noise — –3 dB/octave via Paul Kellet's 7-stage filter. Natural-sounding, effective for speech masking in the 300–3,000 Hz band.
• Brown noise — –6 dB/octave. Deep rumble that blocks traffic, HVAC, and fans. Research suggests low-frequency noise below 500 Hz reduces stress.

Active Noise Cancellation

Three ANC modes run entirely inside AudioWorklet processors for minimum latency:

Phase Inversion Mode

Captures mic audio, inverts its phase through a cascaded 18 dB/octave lowpass filter, and outputs the anti-noise. Simple and immediate, but limited by system latency — effective only for frequencies below roughly 1/(2 × latency). At typical browser latency of 14–40 ms, this works for sub-bass content.

NLMS Adaptive Predictor Mode

The Normalized Least Mean Squares (NLMS) adaptive filter overcomes the latency barrier. It works by learning the periodic structure of ambient noise:

1. Training: The filter uses samples from D samples ago (where D equals the measured system latency) to predict the current sample. The prediction error is used to update the filter weights via the NLMS algorithm: W = W + μ · e(n) · u / (||u||² + ε).
2. Prediction: Using the same learned weights but with current samples, the filter predicts what the noise will sound like D samples in the future.
3. Cancellation: This prediction is inverted and output as anti-noise. By the time it reaches your ears (D samples later), it arrives in phase with the actual noise.

For periodic noise (fans, AC, hums, engine drone), the filter converges within seconds and can effectively cancel the tonal components. A convergence meter shows real-time MSE (mean squared error) — when it drops below 0.01, cancellation is active. The filter length controls how many frequency components can be modelled, while the learning rate controls the speed-stability tradeoff.

Spectral Suppression Mode (Wiener Filter)

The most versatile mode, effective against non-periodic noise like speech, typing, keyboard clicks, and broadband hiss. It uses a 1024-point FFT with 75% overlap (256-sample hops) to analyze the frequency spectrum of incoming audio.

1. Noise estimation: During the first ~0.3 seconds (15 frames), the processor builds a noise power spectrum estimate. After initialization, a Voice Activity Detector (VAD) using energy and spectral flatness analysis determines which frames contain only noise. The noise estimate updates continuously during noise-only frames.
2. Decision-directed Wiener filter: For each frequency bin, the processor computes the a priori SNR using temporal smoothing (α_dd = 0.98). The Wiener gain G(k) = ξ(k) / (ξ(k) + 1) determines how much each frequency is attenuated.
3. Musical noise mitigation: A spectral floor prevents over-suppression artifacts, and 3-point frequency smoothing reduces isolated spectral peaks that cause “musical noise” — the tinkling artifact common in basic spectral subtraction.
4. Overlap-add synthesis: The processed spectrum is inverse-FFT'd and overlap-added with a Hann window for artifact-free reconstruction.

Ultra-Low-Latency Pipeline

The AudioContext uses latencyHint: 0 (numeric zero) to request the absolute minimum latency the platform supports. The mic stream disables echo cancellation, noise suppression, and auto gain control to eliminate processing latency. Mono input (channelCount: 1) halves the processing load. The system measures and displays baseLatency + outputLatency so you can see your actual round-trip performance (Firefox typically achieves ~14 ms, Chrome ~19–40 ms).

A-Weighted Metering & Adaptive Mode

Dual meters show raw dB and A-weighted dB(A) per ISO 61672-1. The adaptive mode uses dB(A) with an attack/release envelope for natural masking volume adjustments.

Privacy

All processing runs in your browser. No audio is recorded, stored, or transmitted. The mic stream and all AudioWorklet processors are terminated when you stop the tool.