Reading Audio Spectrograms: Bands, Noise, and Traces
Describing a recording as muddy or noisy gives an engineer nothing to fix. A spectrogram spreads sound across time, frequency, and energy so the problem has a visible address.
Spectrogram vs waveform
A waveform is an energy envelope over time - one dimension. Loud is tall, quiet is short, but it cannot tell a low-frequency rumble from high-frequency hiss at the same level.
A spectrogram adds the frequency axis. Time runs left to right, frequency runs bottom (low) to top (high), and brightness encodes energy. Voice concentrates between 100 Hz and 4 kHz, sibilance lives above 6 kHz, bass content sits below 100 Hz. With that map in mind, every artifact has a recognizable shape.
What common problems look like
Electrical hum is a set of horizontal bright lines: 50 Hz and its harmonics (100, 150, 200 Hz) running through the entire timeline, fixed in place while content changes around them. Clipping shows up as a fence of vertical bars when voice hits 0 dB - harmonics of a squared-off waveform. No amount of volume adjustment recovers that.
Low bitrates carve the top: audio encoded below MP3 128 kbps or heavily transcoded by platforms flattens into a straight line near 16 kHz with everything above blank. Compare against the original recording and you can count how many times it was squeezed.
Choosing colors and size
For analysis choose viridis: the green-to-blue ramp matches human sensitivity best and separates energy levels most clearly - the de facto standard in engineering reports. inferno and magma (black with warm highlights) pop on dark slides; rainbow looks lively but maps energy to color unevenly, so skip it for serious analysis.
Size follows purpose: 1280x720 reads overall structure fine; step up to 2560x1440 to count harmonics or inspect cutoff edges. The image is a diagnostic, not a deliverable - end with an action: denoise, re-record, or accept.