Audio silence remover with local MP3 export

What this audio silence remover actually cuts

Silence in a recording is not a special marker. It is a run of samples whose level stays below a threshold. This audio silence remover decodes the selected WAV or MP3, measures the largest absolute sample in consecutive 10 millisecond windows, and treats windows below your dBFS setting as quiet. A quiet run becomes a cut only when it lasts at least as long as the minimum silence setting. Short breaths and natural spaces can therefore stay even when long pauses disappear.

The default threshold is −35 dBFS, the minimum pause is 150 milliseconds, and 40 milliseconds of padding remains on each side of every kept section. Those are starting points, not hidden rules. Lower the threshold when quiet words are clipped. Raise it when steady air-conditioner noise prevents pauses from counting as silence. Increase the minimum pause or padding when a conversation sounds rushed after the first pass.

Remove silence from MP3 without uploading it

The original file never goes to a conversion server. The browser reads it in byte slices, Web Audio decodes it in memory, and a background worker finds the audible sections. Only then does the page load its self-hosted MP3 encoder. The worker sends progress based on actual samples encoded, joins the kept sections in their original order, and returns a local blob URL. The player and download button both use that same blob.

This is especially useful for interviews, voice notes, lectures and narration that contain dead air but do not need a full multitrack editor. It is not a noise remover. A fan, hum or room tone above the threshold remains part of the recording, and reducing that noise requires a different signal-processing decision. Use the controls to identify quiet intervals, then listen to the joins before saving the result.

WAV and MP3 input, real MP3 output

WAV and MP3 are accepted because both are broadly decoded across modern browsers. The header is checked instead of trusting the extension, so a PDF renamed to .mp3 is not handed to an audio decoder. An unknown audio container is allowed to reach the browser, because installed codec support can know more than a small signature list. If decoding still fails, use the original WAV or MP3 rather than a file renamed from another format.

The result is always a new 128 kbps MP3. That matters when the input is already MP3: silence cannot be removed from arbitrary compressed frames and joined without decoding, because MP3 frames overlap and carry encoder delay. The kept samples are encoded again, so the output is lossy and not a bit-for-bit copy. Preview it on the receipt instead of assuming that a smaller duration guarantees a good edit.

How the three controls change the edit

Threshold answers “how quiet is quiet?” Minimum silence answers “how long must that quiet level continue before it is worth removing?” Padding answers “how much room should remain around the speech or music?” Keeping these as separate numbers prevents a common failure in one-click tools: treating a soft syllable and a long empty pause as the same thing merely because both contain low samples.

An all-silent file produces no invented audio. Instead, the page asks you to lower the threshold and try again. Mono and stereo are supported. Surround recordings are declined with a specific explanation rather than being downmixed silently, because an automatic downmix can change levels and phase before silence detection has even begun.

Why the MP3 encoder is a separate download

Browsers can decode MP3 but generally do not expose MP3 encoding through MediaRecorder or WebCodecs. This page uses the LAME 3.100 encoder compiled to WebAssembly by wasm-media-encoders. The 132,999-byte WASM file is fetched only after you start a job, from this same origin. It sits at a versioned URL separate from the application bundle, so a compatible rebuild can replace it. The complete LGPL text, the wrapper’s MIT license, exact source commits and build instructions are linked directly below the tool.

The encoder is code, not a service. Loading it does not send the audio anywhere. To check that distinction yourself, follow the local-processing verification guide and watch the Network panel. You should see the page’s own MP3 encoder asset, but no request whose body contains the recording, filename, waveform or settings.

This is intentionally not a video silence remover

Video needs a different chain. Removing two seconds from the soundtrack while leaving the picture untouched creates a file that drifts out of sync. A truthful video tool must demux the container, cut both tracks on matching timestamps, re-encode them and mux a playable result. None of that is implied by this audio page, and video files are not accepted. If you need a still view of amplitude before editing, the audio waveform generator exports measured peaks as SVG or PNG. If your source is an animation with no sound, the GIF-to-MP4 converter preserves its real frame timing.

Large recordings use memory, not an upload quota

Compressed audio expands to floating-point samples during decoding, so a long recording can occupy much more memory than its file size suggests. The page shows the shared large-file advice before work begins, reads file bytes in slices, copies decoded channels in bounded pieces and reports real progress. It never rejects a recording merely because of byte count. On a memory-constrained phone, close other tabs first and use Cancel if the job is not worth the device load.