Audio Denoiser
Choose a short part that contains only the unwanted steady sound, such as fan hum or hiss. The tool averages its frequency spectrum, attenuates matching energy across the file, and rebuilds the waveform with overlap-add. Listen to the unchanged original and the reduced version at the same level before saving WAV. This is a conservative spectral subtraction tool for roughly stationary noise, not voice isolation or source separation. Strong settings or an incorrect noise sample can produce metallic artifacts and damage speech.
Key features
- Build a per-channel noise magnitude profile from a user-selected noise-only interval of at least 0.25 seconds.
- Apply a 2048-point STFT with a 512-sample hop, a spectral floor, three-bin and temporal gain smoothing, and normalized overlap-add.
- Set strength from 0 to 2.5; zero returns the exact decoded samples before WAV quantization.
- A/B play the unchanged decoded source and reduced result without loudness normalization.
- Compare RMS dBFS inside the selected noise interval, whole-file RMS and peak; download 16-bit WAV and a settings/metrics JSON manifest.
How to use
- Choose a browser-decodable local file no larger than 8 MiB, or load the generated hiss-and-tone WAV.
- Set the start and end of a segment containing noise alone; avoid speech or music in the sample.
- Run reduction, inspect the noise-interval RMS change and listen to original and reduced versions.
- Adjust strength if speech sounds thin or metallic, then save the WAV and optional JSON report.
Use cases
- Reduce constant fan or air-conditioner hiss under a voice recording.
- Compare two strength settings using an unchanged-level A/B playback.
- Record the selected noise interval and objective before/after values for a later editing session.
Frequently asked questions
How does the noise reduction work?
It averages STFT magnitudes from complete frames within the selected noise-only interval separately for each channel. For every frame it subtracts a strength-scaled estimate from the observed magnitude, leaves the phase in place, smooths neighboring and consecutive gains, and uses normalized overlap-add to reconstruct audio. The idea follows spectral subtraction; the exact settings here are documented in the JSON report.
Can it isolate a voice from music or changing background sounds?
No. It expects the unwanted sound to be approximately stationary and represented in the selected interval. Sudden impacts, overlapping music and moving sources are not separated. If the profile includes voice, the same frequencies can be removed from the voice elsewhere.
Does a larger dB reduction prove better sound?
No. The reported reduction is RMS change only in the selected noise interval. It is not signal-to-noise ratio, speech intelligibility or perceived quality. Listen to both versions for watery or metallic artifacts and lost consonants.
What do strength 0 and the output format mean?
Strength 0 copies the decoded floating samples exactly in memory, but the saved file is always 16-bit PCM WAV. A lossy or higher-bit-depth source may change precision, and samples outside ±1 are clipped in WAV; the report counts them.
Why does it sound different from another editor?
The browser decodes and may resample the input to 48 kHz. This implementation uses a fixed 2048-point spectrum, 512-sample hop, 6% spectral floor and fixed gain smoothing; another decoder or reduction method may differ. The report includes settings and source SHA-256.
Is my audio uploaded?
No. The file is decoded, profiled, processed and played locally in your browser. Downloaded JSON has metadata and measurements but no audio samples; the WAV is saved only when you request it.
Privacy
Selected audio and its noise profile remain in browser memory. The site does not upload PCM or send it to an analysis service. The optional JSON includes the file name, byte size and SHA-256, interval, settings and metrics, but no audio samples.
Comments & questions