Audio Noise Reduction

The Audio Noise Reduction tool applies a high-pass filter that weakens everything below a chosen frequency. That reduces background noise concentrated in the low end, such as electrical hum, fan rumble, air-conditioning drone, and low room vibration. It is deliberately simple, and the red notice at the top of the page says so: this is basic filtering, not an AI noise-reduction algorithm that can tell speech from noise. It can do almost nothing about noise that shares the voice's frequency range (street chatter, other people talking, keyboard clicks, high hiss), and it cannot fix clipping or echo. Treat it as a light tool for removing low-frequency rumble.

How to use it

1. Click the dashed 'Click to upload audio' box and choose an audio file. Your browser has to be able to decode it. The box shows the file name.
2. Drag the 'Noise reduction strength' slider. It has three positions: 'Light' on the left, 'Medium' in the middle, and 'Strong (may affect bass tone)' on the right. It starts at the middle, and the current cutoff is shown under the slider. Under 'Output format' choose 'WAV 16-bit (default, most compatible)', 'WAV 32-bit float (no clipping or quantization, larger file)' or 'MP3 128 kbps (small file, lossy)'.
3. Click 'Start Noise Reduction'. The button shows 'Processing…'. This tool does not download any extra processing core.
4. When it finishes, a player appears. Listen, and if it sounds right click 'Download Result'. The saved file is named denoised.wav, or denoised.mp3 if you chose MP3.

What the three strengths actually do

The filter's cutoff is one of three fixed values: 80 Hz for 'Light', 150 Hz for 'Medium' and 250 Hz for 'Strong'. (Earlier versions used 80, 200 and 320 Hz; the 200 Hz default could cut into the fundamentals of some male voices, so the values were lowered.) Sound below the cutoff is progressively attenuated. So 'Light' removes little beyond the deepest rumble, while 'Strong' also cuts into some of the bass. A usage scenario: you recorded a talk on your phone in a quiet room and hear a steady low hum on playback. Try 'Medium' first; if the hum is still there, try 'Strong' while listening for whether the voice has become thin. Note that a large part of an adult male speaking voice sits around 100 to 200 Hz, so even after the reduction the 250 Hz of 'Strong' can make a voice sound thinner, and it is not wise to push the slider to the maximum by default. How it sounds varies from recording to recording, and we have no effectiveness figures to quote.

Supported input and known limits

• There is a single high-pass filter and nothing else: no noise sampling, spectrum analysis, or adaptive processing, so noise outside the low frequencies is essentially untouched.
• Only three strengths are offered (cutoffs of 80, 150, and 250 Hz). You cannot set your own frequency.
• There are three output formats, whatever you upload: WAV 16-bit PCM (default, most compatible), WAV 32-bit float (no clipping or quantization, about twice the size) and MP3 128 kbps (constant bit rate, lossy; about 0.96 MB per minute, versus about 10.6 MB per minute for 16-bit stereo 44.1 kHz WAV). MP3 is encoded in the browser by the bundled lamejs and is only loaded the first time you process as MP3. Mono and stereo are supported; with more than 2 channels only the first two are kept, and sample rates MP3 cannot use (such as 96 kHz) are resampled to 48 kHz first. MP3 adds another lossy compression step to the processed sound; choose WAV if you want to avoid that.
• The browser must be able to decode the upload itself, and supported formats differ between browsers. If decoding fails you will see 'Processing failed. Please make sure you uploaded a valid audio file.'
• The entire audio is decoded to uncompressed data in memory, so long recordings use a lot of memory and may be slow or fail.
• Channel count and sample rate are kept, but loudness is not re-leveled afterwards, so cutting low frequencies can change the perceived volume a little.
• Each run creates a new browser audio context. If you process many files in quick succession and the page gets sluggish, refresh it.
• There is no before-and-after comparison control; to compare, play the original file and the result separately.

FAQ: Why is there still noise after processing?

Because the tool only affects low frequencies. If the noise is hiss, keyboard clatter, or crowd chatter, it overlaps the voice range and a high-pass filter cannot remove it. That calls for dedicated noise-reduction software, or avoiding the noise at recording time.

FAQ: The result sounds muffled or thin. What now?

The cutoff is probably too high and is cutting into the real low-frequency content of the sound. Go back to 'Light' or 'Medium' and run it again. Each run starts from the file you uploaded, so effects do not stack.

FAQ: Why is the default a WAV?

By default the processed sound is written straight to an uncompressed WAV (16-bit or 32-bit float). That needs no extra encoder and avoids another lossy compression step, at the cost of file size. If you need a small file, pick 'MP3 128 kbps' under 'Output format'.

FAQ: Can I clean up the audio in a video?

The upload box accepts audio files only. Extract the audio from the video with another tool first, then process that.

Privacy and processing

Everything happens in your browser. The file is read into memory, decoded with the browser's built-in audio API, filtered, and written out as a WAV (or encoded to MP3 by the bundled lamejs if you choose that). Looking at the code, the tool makes no request that uploads audio to a server and does not download any processing core from outside. The result is played and downloaded through a temporary local link that is released when you change files or leave the page.