Split Audio at Silence

How the Silence Splitter works

Split audio at silence: what it does and why

Silence splitting is the automatic version of the region exporter: instead of typing where the cuts go, the tool finds them. It measures the loudness of the file in ten-millisecond windows, marks a window as quiet when its RMS level is below the threshold, and treats any run of quiet windows longer than the gap setting as a boundary. Sound between boundaries is a clip, provided it lasts at least the minimum clip length.

Why split on silence

Most long recordings are really a string of short ones with pauses between them. A voice actor records 200 game lines in a single take with a beat of silence after each; a vinyl side digitised as one file holds five songs separated by the run-out groove; a sample-pack session is a hundred drum hits in a row. Cutting those by hand in an editor means zooming, scrubbing and exporting a region at a time — at 20 seconds per cut, a 200-line session is over an hour of clicking. Splitting on silence does the same job in one pass, with every cut placed at the quietest point rather than wherever the mouse landed.

Where it gets used

• Game and app audio: one file per voice line or UI sound, ready to drop into an engine or an audio sprite.
• Digitising records and tapes: one file per track from a whole LP side.
• Language and dictation material: one sentence per file for flashcard apps such as Anki.
• Sample libraries: separate one-shots from a recorded run of hits.
• Speech datasets: short utterances for training or testing speech recognition.

The settings

Threshold. Levels are in dBFS, decibels below digital full scale. A studio recording with a clean noise floor splits well at -50 or -60; a room recording, a phone memo or a vinyl rip with surface noise needs -35 to -40 so the noise floor itself is not counted as sound. If you get one enormous clip, the threshold is too low for the noise in the file; if you get dozens of tiny ones, it is too high and pauses inside the material are being treated as gaps.

Gap length. This is what separates a pause from a boundary. Half a second suits voice lines and sound effects; two or three seconds is safer for songs, where a quiet intro or a breath between verses should not start a new file. Natural speech has pauses of 0.2 to 0.5 seconds between phrases, so a gap below about 0.4 seconds will start cutting sentences apart.

Minimum clip length drops clicks, coughs and stray noise that would otherwise become one-second files of their own. Padding adds a little of the surrounding silence back on to each clip so the cuts do not feel clipped; fifty milliseconds is enough to keep a soft attack intact without leaving obvious gaps.

Tips and pitfalls

• Measure the noise floor first. The silence finder shows every gap with its length at a given threshold, so you can try a level and see whether the gaps it finds match the pauses you expect before splitting.
• Leave the pause when recording. A clean second of silence between takes makes splitting reliable; talking straight into the next line does not.
• Fades and reverb tails are quiet but not silent. A song that fades out, or a sound effect with a long tail, may be cut short. Lower the threshold or raise the padding.
• Classical and live albums often have near-silent passages inside a movement. Use a long gap, three to five seconds, or split by hand.
• Want one file with the pauses removed? That is the silence remover; the silence trimmer only trims the start and end. For tidying up a spoken recording, see how to clean up a voice recording.

How to read the result

The stats open with the number of clips and the rule that produced them, then Kept: the total length of the clips, the length of the original and the percentage kept. On voice work with short pauses expect 70–90%; much lower means the threshold is eating quiet speech. The first ten clips are listed with their start and end times and duration, and a final line counts any further clips in the zip. Check that the count matches what you expected — the number of songs on the side, or the lines in the script — before downloading.

The exported files keep the source channel count and sample rate and are numbered in order, name-001, name-002 and so on, in the format selected below the settings. Nothing is normalised or otherwise processed; use the audio normalizer if levels need matching afterwards.

Related tools