Volume and loudness: how to boost audio safely

Volume, gain and decibels

In a digital file, the waveform is a sequence of samples with finite values. Gain multiplies every sample by a factor: doubling the amplitude is a gain of 2, and in decibels that is about +6 dB. Reducing the amplitude to half is -6 dB, and ten times the amplitude is +20 dB. The decibel scale is logarithmic because human hearing responds to ratios rather than absolute differences.

Perceived loudness is a different measurement. Two signals with the same peak amplitude can seem very different in loudness, because hearing is more sensitive to some frequencies than others and because sustained sounds seem louder than short ones. As a rough rule of thumb, a 10 dB increase is heard as roughly twice as loud, but that is a perceptual convention, not a guarantee for every recording.

Clipping and distortion

Digital audio has a defined maximum sample value, usually called 0 dBFS. A sample that would be louder than that has nowhere to go, so it is cut off at the ceiling, which is hard clipping. Clipped waveforms develop flat tops and sharp edges, and those edges add harmonics that were never in the recording; the result sounds harsh, fizzy or crackly, and it is permanent.

There is also a subtler danger between samples. True peak meters estimate the analog waveform that a digital-to-analog converter will reconstruct, and it can overshoot the sample values. That is why a file whose samples peak at 0 dBFS can still distort on playback, and why leaving a little headroom, often a decibel or two below the ceiling, is safer than normalising to the absolute maximum.

Normalization is not the same as amplification

Amplification applies a fixed gain with no analysis: every sample is multiplied by the same number, so the relationship between quiet and loud passages is unchanged. Normalization measures something first and then chooses a gain. Peak normalization finds the loudest sample and raises the file so that it sits just below the ceiling, which makes the recording as loud as it can be without clipping.

Loudness normalization targets perceived level instead, using a measurement based on how hearing works, and it may boost or attenuate the file to reach a target. That distinction matters for delivery: streaming services and broadcasters measure loudness on their own terms and turn files up or down to a house level. A podcast pushed far above the usual range does not end up louder on those platforms; it simply gets turned down, with less headroom left for its own peaks.

Loudness targets are usually expressed in LUFS and are based on standards such as EBU R128 and ITU-R BS.1770. The exact target differs between platforms and broadcasts, and the numbers are revised over time, so check the current delivery spec for the destination rather than relying on a remembered figure.

A practical workflow for a quiet recording

Start by deciding whether the whole file is quiet or only parts of it. Uniformly low speech, such as a lecture recorded from the back of a room, is a candidate for straightforward gain. A podcast with quiet and loud guests needs leveling or compression rather than one global multiplier, because a boost that fixes the quiet speaker will push the loud one into distortion.

Then check the peaks before you raise anything. If the loudest peak currently reaches -20 dBFS, there is roughly 20 dB of headroom to use, so even a loud-sounding boost can be safe; if peaks already sit near -1 dBFS, the file is not actually quiet. Apply the gain, leave one to two decibels of headroom, and listen to the result at a realistic level rather than on a phone speaker turned down.

• Read the file's peak and loudness readings before touching the level.
• Raise quiet, uniform recordings in one pass rather than stacking boosts.
• Leave a decibel or two of headroom instead of pushing peaks to 0 dBFS.
• Export an uncompressed WAV while editing, then encode the delivery format once.
• Re-check the loudness of the finished file against the destination's spec.

How speed changes interact with pitch and duration

Changing playback speed is not the same as changing gain. Speeding a file up shortens its duration, and a naive resample also raises the pitch, which makes voices sound like chipmunks. Time-stretching avoids that by changing tempo without resampling the waveform, which is what FFmpeg's atempo filter does and what a pitch-preserving speed changer exposes.

Duration follows a simple relationship: the new length is the original divided by the speed. A three-minute recording at 1.5x becomes two minutes, and at 0.5x it becomes six minutes. Pitch stays the same throughout, but extreme speeds stretch or squash the spaces between transients, so percussion and complex music can smear. When you combine a speed change and a gain boost, do both deliberately and check peaks again, because a stretched file can contain new peaks that were not obvious in the original.