Audio Tools

Peak normalisation raises a file until its loudest single sample touches the ceiling. That sounds like it should work, and it does not: one stray drum hit or door slam can be the peak, leaving an otherwise quiet recording untouched. Loudness normalisation measures perceived level over time in LUFS instead, which is why two files normalised to the same LUFS value actually sound equally loud.

Useful targets: around -14 LUFS integrated for music going to Spotify or YouTube, -16 LUFS for stereo podcast audio to match Apple’s recommendation, and a true-peak ceiling of -1 dBTP so that lossy encoding does not push samples into clipping after conversion. Streaming platforms turn anything louder back down anyway, so mastering at -9 LUFS buys you nothing except lost dynamics.

MP3 stores audio in frames of 1152 samples — about 26 milliseconds at 44.1 kHz — and a cut can only land on a frame boundary without re-encoding. This is why a trim occasionally leaves a few milliseconds you did not ask for. Encoders also add padding at the start and end, which is what produces the small gap between merged tracks that should have been continuous.

Cutting in the middle of a waveform leaves a sudden jump in amplitude, and a sudden jump is a click. A 5–10 millisecond fade at each edit point removes it without being audible as a fade. When merging files recorded on different days or different devices, normalise after merging rather than before, otherwise you are matching each file to a target instead of matching them to each other.

Changing playback rate by resampling moves speed and pitch together — this is the tape effect, and at 0.8x it drops the pitch by roughly four semitones. Time-stretching uses a phase vocoder to change duration while holding pitch constant, which is what you want for a podcast played at 1.25x. It costs something: transients smear, and sharp percussion picks up a faint metallic ring past about 1.5x or below 0.7x.

The slowed + reverb sound deliberately uses the first method. The track is resampled to roughly 0.75–0.85x so pitch drops with the tempo, then a long reverb with a decay in the 2–4 second range is mixed in behind it. Push the rate below 0.7x and the low end turns muddy; keep the reverb mix under about 35% or the vocal stops being intelligible.

Silence detection works on two numbers: a threshold and a minimum duration. A threshold near -40 dBFS catches room tone in a decently treated space, but a noisy room may need -30, and setting it too aggressively clips the quiet tails off words. Requiring 300–500 milliseconds before a gap counts as silence stops the tool from eating natural pauses between sentences, and leaving 100–200 milliseconds of padding at each edge keeps speech from sounding clipped.

For interviews, lectures and voice memos recorded on a single microphone, converting to mono halves the file size and loses nothing, since both channels carry the same signal. Be careful with material recorded on two mics or processed with stereo widening: summing to mono can cancel out-of-phase content and make parts of the mix disappear. Check by listening to the mono version before you delete the stereo original.