Audio formats, bitrates and why one track is louder

Keep the master lossless, WAV while you work and FLAC to store it. Deliver in AAC or Opus where you know what will play the file, MP3 where you do not. Speech goes out in mono at a low bitrate. Set the level by LUFS, not by peak, and leave a decibel of headroom under full scale.

That is the decision. The rest explains why.

What each format is for

Lossless means the decoded samples are identical to what went in; FLAC saves space by predicting each sample from the previous ones and storing only the error. Lossy formats discard information permanently, using a model of what the ear will not notice.

uncompressed   44100 x 16 x 2 / 8 = 176400 B/s = 10.6 MB per minute
lossy          192000 / 8 x 60    = 1.44 MB per minute

Sample rate and bit depth

Sample rate is how often the waveform is measured. The Nyquist limit says a sampled signal carries frequencies up to half the sample rate and nothing above, so 44.1 kHz covers everything below 22.05 kHz, and human hearing stops around 20 kHz, earlier with age. Use 48 kHz alongside video; higher rates only help during production.

Bit depth sets dynamic range, not detail. Each bit is worth about 6 dB, so 16 bit puts the noise floor roughly 96 dB below full scale and 24 bit roughly 144 dB, of which converters deliver about 120. Those bits are recording headroom: keep the level conservative, then raise it later without lifting audible hiss. Finished, levelled files do not need them.

Upsampling gains nothing: interpolating 44.1 kHz to 96 kHz doubles the file without adding information, and padding to 24 bit brings the original noise floor with it. Every resample runs a filter, so a trip to 48 kHz and back is a small loss for nothing.

Bitrate, and where more stops helping

Figures are kbps. MP3 shows its age at both ends: below about 96 kbps it smears audibly, and above about 192 kbps improvements get very hard to hear, so 320 kbps buys certainty rather than quality. AAC needs around two thirds the bitrate for the same result. Opus does better again, and its lead widens as the bitrate drops, because at low bandwidths it switches to a mode built around how speech is produced, so voice needing 64 kbps as MP3 is fine at half that.

Raising the bitrate when converting one lossy file to another restores nothing. A 128 kbps MP3 re-encoded at 320 kbps preserves the artefacts of the small file faithfully in a large one, and the second encoder spends bits treating them as real audio. Work from the lossless master.

Mono, stereo and folding down

Stereo costs about twice as much as mono at the same quality. One person at one microphone produces no stereo information, so a spoken word file in stereo is two near identical channels and mono halves the size for nothing lost. Music and room recordings stay stereo.

Folding down is not free. The downmix averages the channels, so anything out of phase between left and right cancels: a widening effect or a reversed microphone lead can thin out or vanish in mono only. Check it before publishing, and attenuate by about 3 dB, since summing raises the level.

Peak, RMS and LUFS

Peak is the highest sample value, in dBFS, with 0 as the hard digital ceiling. Normalising to peak scales the file so its loudest sample lands on the target, so a quiet recording with one door slam in it stays quiet: the slam moves to the ceiling, everything else barely moves. RMS averages energy over a window, closer to what is heard but treating 60 Hz and 3 kHz as equal. LUFS weights the measurement to model hearing, gates out quiet passages, and reports one figure for the whole programme. Platforms measure LUFS.

Because platforms pull everything towards a target, mastering louder no longer plays louder; the platform turns the track down and the listener gets the squashed dynamics without the loudness. That is how the loudness war ended, after two decades of harder limiting left masters with almost no gap between peak and average level, heard as flat and tiring.

Clipping is a sample that would exceed full scale being truncated to the maximum, flattening the tops of the waveform and generating harmonic distortion. True peak goes further: the waveform reconstructed between samples can rise above the highest stored sample, so a file measuring -0.1 dBFS can hit +0.5 dBTP and clip on the way out. Leave the ceiling at -1 dBTP, lower if a lossy encode follows. Audio Volume can normalise to a target instead of shifting by fixed decibels.

Clicks, joins and trims

Cut a waveform at an arbitrary point and it will usually be part way through a cycle, at a non zero value, so the output drops to silence in one sample. That step is a broadband impulse, heard as a click. Five to twenty milliseconds of fade at each edit point removes it and is too short to hear; at a join, a crossfade of ten to fifty milliseconds is cleaner still.

Joins need agreement. The files must share a sample rate, because a 48 kHz file concatenated into a 44.1 kHz stream plays about nine percent slow and low, and a channel count, or a section lands in the wrong channels. Match loudness too. Resample with Audio Converter, level with Audio Volume, then use Merge Audio. Audio Speed handles deliberate speed changes: resampling moves pitch and speed together, while holding the pitch needs time stretching, which can smear transients.

Trimming can skip re-encoding entirely, at no cost in quality. WAV is sample accurate, since the trim is a byte copy; compressed formats are built from frames, and a copied cut can only land on a frame boundary.

MP3 frame = 1152 samples = 26.12 ms at 44.1 kHz
cut asked for at 12.000 s
nearest boundaries: frame 459 = 11.990 s, frame 460 = 12.016 s

AAC uses 1024 sample frames, about 21 ms at 48 kHz, and Opus normally uses 20 ms packets. MP3 also has a bit reservoir, letting a frame borrow space from earlier ones, so the first frame after a cut can be missing data and glitch. Where the boundary must be exact, let Trim Audio re-encode and accept one generation of loss.

What a browser can chew through

All of this happens in memory. Files are decoded to raw samples before anything can be measured, faded, joined or re-encoded, usually as 32 bit floats: an hour of stereo at 44.1 kHz is about 635 MB as 16 bit PCM and 1.27 GB decoded. A tab's memory ceiling sits far below what a desktop application can address, so the limit is length, not format. Trim first, or split a long recording and merge the parts.