Remove static hiss from audio inside browser memory
Cleaning fan hiss, microphone self-noise, and office hum out of sensitive voice recordings should never require uploading private interviews to remote cloud queues. Decoding tracks into floating-point arrays inside browser RAM lets a dedicated worker attenuate stationary background noise without network calls.
Measuring background noise floor with 20th-percentile blocks
Dropping an audio track onto Drop a noisy audio track here or clicking Load sample noisy clip decodes waveforms into linear PCM arrays shaped like the MDN Web Audio AudioBuffer interface. Before applying filters, the worker partitions your file into 30 ms blocks and measures RMS level in each window.
Sorting those 30 ms RMS values and reading the 20th percentile gives an accurate noise floor estimate in dBFS. Because spoken syllables occupy louder percentiles, tracking p20 measures background hiss cleanly rather than vocal peaks. You can also test microphone input levels and room noise before recording to capture cleaner takes upfront.
Gain-staging loud hiss into the adaptive noise expander
After filtering rumble and hum, our audio noise remover evaluates residual noise across 4,096-sample blocks inside everyday.worker.ts. When post-filter floor RMS exceeds 0.0032 (around -50 dBFS), gain staging temporarily scales loud hiss down into the adaptive expander’s sweet spot before restoring level afterward.
An adaptive 90 Hz sidechain detector follows voice envelopes with a 3 ms attack and 50 ms release while tracking room tone over 40 ms down and 2.2 seconds up. That separation prevents quiet breaths from chopping abruptly while clamping steady static during silent gaps.
Why do notch and high-pass filters strip 50/60 Hz hum?
Alternating-current ground loops inject sharp tonal spikes at 50 Hz or 60 Hz plus upper harmonics, while desk vibration rumbles below vocal fundamentals. Cascading narrow Q=22 biquad notches after a sub-bass high-pass filter removes electrical buzz and rumble before the expander tracks speech cleanly.
Cutting desk rumble with the 20 Hz to 200 Hz high-pass slider
Setting Sub-bass rumble high-pass (20 Hz to 200 Hz, defaulting to 80 Hz) runs a second-order biquad filter patterned after the MDN BiquadFilterNode filter documentation. Removing sub-80 Hz air-conditioner vibration prevents heavy bass energy from falsely opening the noise gate between spoken words.
Keep the cutoff near 70 Hz to 90 Hz for adult narration, or raise it toward 120 Hz when cleaning handheld phone memos recorded outdoors in wind. Once cleaned, dry voice tracks blend cleanly under music beds inside Audio Mixer without low-end mud.
Eliminating AC mains harmonics with Q=22 biquad notches
Power supply hum rarely sits strictly at one frequency; nonlinear microphone preamps add strong second and third harmonics at 120 Hz/180 Hz or 100 Hz/150 Hz. Selecting 60 Hz or 50 Hz under AC mains hum notch inserts three cascaded Q=22 notches targeting all three spikes whenever they fall below Nyquist limits.
Because a Q factor of 22 creates an extremely narrow cut, neighboring vocal warmth remains untouched. Leaving AC mains hum notch set to Off bypasses those three biquads completely when cleaning pure battery-powered field recordings.
Noise reduction strength, hum notch, and filter presets
Comparing Gentle, Standard, and Aggressive presets alongside AC mains hum notch and Sub-bass rumble high-pass helps balance hiss suppression against vocal naturalness. Filter transfer math matches standard second-order coefficients defined in the W3C Web Audio API 1.1 specification.
Balancing Gentle mode room preservation versus Aggressive gating prevents clipped consonant tails while stripping stationary hiss. Choosing Gentle leaves subtle studio air intact, whereas Aggressive clamps loud laptop fan noise between sentences.
| Control | Available options or range | Default setting | Internal DSP behavior | Best recording scenario |
|---|---|---|---|---|
| Gentle | Radio button option | Unchecked | Blends 62% gated audio with 38% filtered signal |
Studio mics needing subtle hiss reduction |
| Standard | Radio button option | Selected | 2.6x SNR gate margin, 0.12 floor gain |
Podcasts, calls, and everyday voice memos |
| Aggressive | Radio button option | Unchecked | 3.6x SNR margin, 0.04 floor + -4.5 dB 11 kHz shelf |
Loud laptop fans and cheap headset hiss |
| AC mains hum notch | Off, 50 Hz, 60 Hz | 60 Hz | Three Q=22 notches at 1x, 2x, 3x frequency |
Buzzing USB interfaces or AC wall hum |
| Sub-bass rumble high-pass | 20 Hz to 200 Hz (5 Hz step) |
80 Hz |
Second-order biquad HPF (Q = 0.707) |
Desk bumps, footsteps, and HVAC rumble |
Once broadband hiss drops cleanly, quiet gaps between segments become easier to detect automatically. Exporting the cleaned WAV and running silence-based podcast chapter splitting yields accurate segment boundaries without false triggers from background noise.
Tip: Click Load sample noisy clip to hear how Standard mode paired with 60 Hz notch filtering drops the
-28.1 dBFShiss and hum floor cleanly in one pass before processing your own recordings.
Compare original and cleaned WAV audio side by side
Every adjustment updates the 160-bar SVG Cleaned audio waveform and displays exact Noise floor changes from input dBFS to output dBFS alongside a dB quieter badge. Auditioning changes directly in browser memory avoids exporting dozens of trial files.
Using synchronized A/B playback and noise-floor telemetry
Clicking Original audio or Cleaned audio under A/B playback comparison pauses the inactive player, copies the exact currentTime offset, and continues playback on the selected track. Hearing the same spoken consonant back-to-back catches over-gating or hollow filtering immediately.
For compressed sharing after verification, WebCodecs browsers offer Download .webm encoding aligned with the IETF RFC 6716 Opus Audio Codec standard, alongside Download .m4a (128 kbps AAC). Universal Download cleaned WAV writes uncompressed 16 bits per sample PCM (encodeWav) with exact byte sizes displayed.
Applying makeup gain with soft-knee ceiling protection
If cutting rumble or noisy peaks leaves a recording quiet, raising Makeup gain (-3 dB to +12 dB in 0.5 dB steps) boosts final output cleanly. Every boosted sample passes through softLimitCeiling clamped at -1.0 dBFS (85% linear knee up to -1.0 dBFS), preventing harsh integer clipping.
Creators archiving finished podcast masters or checking whether a noisy MP3 came from a low-bitrate source can verify spectral frequency cutoffs and update track tags right after exporting. All temporary object URLs stay strictly inside your active tab until closed.