Stereo balance

The balance column reports the whole-track RMS level difference between the two decoded stereo channels. It is descriptive: panning, microphone placement, mastering decisions and DC offset can all create a non-zero result without a recording defect.

Calculation#

The analyzer accumulates unweighted energy for both channels. Because RMS is the square root of energy divided by the same frame count, the difference is calculated as:

right minus left = 10 × log10(E_R / E_L) dB

This follows the ordinary RMS definition. In code, the two logarithms are subtracted instead of dividing the energies first, so very unequal finite energies cannot overflow the ratio. Common gain or reversing either channel's polarity leaves balance unchanged; swapping channels reverses its sign. There is no minimum duration or absolute-level gate, so even a short, quiet, nonzero float signal has a measurement.

A positive result means the right channel is louder; a negative result means the left channel is louder. The display uses the more readable form R +x.x dB or L +x.x dB. If every sample in one channel is zero, the result is L silent or R silent; an infinite dB value is never exported. Both channels silent, invalid values and layouts other than exactly two channels have no reading.

CSV keeps the unambiguous signed form in balance_right_minus_left_db, plus balance_silent_channel for the two silent states.

Limits#

This is a broad-band, unweighted, whole-track RMS measurement. It does not model human loudness, does not establish channel assignment, and does not distinguish an intentional asymmetric mix from a wiring problem. A short loud event can dominate energy, while a frequency-specific imbalance may not be obvious in the single number.

DC is deliberately included: mean square = variance + mean². Equal AC energy with different constant channel biases can therefore give a nonzero balance. That is raw signal balance, not an AC-only level comparison. Consult DC offset separately; removing DC here would silently change the quantity reported. Likewise, replacing it with a loudness difference would require a separate definition of channel weighting, duration and gating, as distinct from ITU-R BS.1770-5 programme loudness.

Equal RMS does not imply identical channels, equal perceived loudness, or mono cancellation. Those questions need the waveform relationship, frequency content and the phase measurements. This column has no good/bad balance threshold. Accumulation requires constant memory and one energy sum per channel in the existing decode pass; the two logarithms are evaluated once at the end.

Tests and manual fixtures#

cargo test -p flaccompagnon-core analysis::stereo
cargo test -p flaccompagnon-core --test stereo_balance --test stereo_validity

Analytical unit tests cover the amplitude/energy conversion, common gain, huge energy ratios, polarity-independent energy, raw DC and exact silence. Float WAV integration fixtures cover sixteen-frame excerpts from very quiet to above full scale, known constant biases and silent-channel states. Malformed-frame injection verifies that a valid prefix is withheld after invalid stereo input.

ffmpeg -n -f lavfi -i 'aevalsrc=0.1*sin(2*PI*1000*t)|0.05*sin(2*PI*1000*t):s=48000:d=5' -c:a pcm_s24le balance-left.wav
ffmpeg -n -f lavfi -i 'aevalsrc=0.1*sin(2*PI*1000*t)|0:s=48000:d=5' -c:a pcm_s24le right-silent.wav

balance-left.wav should show L +6.0 dB (the exact amplitude-ratio result is 6.0206 dB). right-silent.wav should show R silent and no L/R correlation. For an independent numerical check, run ffmpeg -i balance-left.wav -af astats -f null - and compare the two channel RMS levels.