DC offset

The DC (%) column measures the average displacement of decoded audio from zero. It shows the largest absolute channel mean as a percentage of full scale, with three decimal places. Hovering shows each channel's signed mean, in decoded channel order. For example, channel means of +1% and −2% produce a cell value of 2.000.

Calculation#

For each channel, using every decoded sample in the file:

mean[channel] = sum(samples[channel]) / frame_count
max_abs       = max(abs(mean[channel]))
display       = 100 × max_abs

Samples use normalized amplitude: +1.0 is positive full scale, so 0.01 is 1%. The calculation includes silence and the final frame. There is no frequency weighting, gate, resampling, filtering or minimum duration. Each channel has its own compensated double-precision sum to preserve small contributions when large positive and negative samples cancel. Opposite offsets in different channels never cancel through a mono downmix.

This follows the arithmetic-mean interpretation of DC offset described in the Audacity manual and FFmpeg's DC correction documentation. A persistent displacement consumes headroom on one side of the waveform. The summation uses Neumaier's correction; it improves floating-point accuracy without making the calculation exact for all possible inputs.

Availability and interpretation#

• Mono, stereo and multichannel streams with 1–32 decoded channels are supported, independently of sample rate.
• Digital silence measures zero; an empty stream has no reading. Constant nonzero samples are measured even without an audible alternating signal.
• Any malformed frame or non-finite sample withholds the entire reading. Failed or skipped decoding and older reports show —.
• For DSD decoded with FFmpeg, the value describes the resulting PCM stream and may depend on the conversion filters. Header-only DSD analysis has no DC reading.
• Very short excerpts, incomplete low-frequency periods and asymmetric edits can have a nonzero mean. Unequal positive and negative peaks alone do not imply a nonzero mean.
• A whole-file mean does not locate changes over time. Equal positive and negative bias in different parts of a track can cancel; silence can dilute a localized offset.
• 0.000 is rounded display precision, not proof of exact zero. Signed tooltip values suppress negative zero after rounding.
• Float audio outside the nominal ±1 range is measured without clipping or clamping, so a result can exceed 100%.

This is a descriptive measurement with no automatic warning threshold. It does not establish hardware failure, audibility, or lossless provenance. It does not change the three authenticity verdicts or modify audio files.

For example, a linear fade from 0.5 to zero has a whole-file mean of 0.25, even though it is changing throughout. Two equally long sections with opposite offsets have zero overall mean. These are correct arithmetic means and require context before interpreting them as a persistent bias.

Cost and alternatives#

The calculation visits each decoded sample once and keeps two double-precision values per channel, independent of file duration. It shares the numerical summation helper with the discontinuity detectors; the measurement does not buffer the track.

A median or trimmed mean would answer a different question and could hide the contribution of short biased sections. A windowed mean could locate changes, but would need a window duration, overlap and additional report fields. A high-pass filter would modify the signal being measured. These alternatives are not implemented, and none is needed to define the current whole-file arithmetic mean.

Reports#

JSON stores dc_offset.channel_means and dc_offset.max_abs in normalized amplitude, without display rounding. Older JSON files without dc_offset still load and need reanalysis to obtain the measurement.

CSV exports dc_offset_max_abs and dc_offset_channel_means, also in normalized amplitude rather than percent. Channel means are separated by semicolons in decoded order, for example 0.01;-0.02. Both fields are blank when unavailable; a measured zero is exported as 0.

Automated verification#

cargo test -p flaccompagnon-core dc_offset
cargo test -p flaccompagnon-core --test dc_offset -- --ignored
node --test tests/analysis-cells.test.mjs tests/search.test.mjs

The ordinary tests inject exact integer or binary-fraction offsets into balanced signals. They check channel independence, silence, asymmetric zero-mean waves, partial periods, fades, opposing sections, invalid tails, large float cancellation, frame-count overflow, single-frame float WAV values above full scale, mono/stereo/multichannel WAV decoding, JSON compatibility and CSV units. UI tests cover signed tooltips, percentage conversion, numerical sorting, search and missing values.

The optional reference test uses the installed FFmpeg binary to compare per-channel astats means and independently encode FLAC. It requires FFmpeg on PATH; no third-party implementation source is used. The comparison allows the half-unit rounding error of FFmpeg's six-decimal output. FLAC decoding must preserve the same channel means as the original integer WAV.

Manual test#

Create a five-second stereo tone with +1% on the left and −2% on the right:

ffmpeg -n -f lavfi -i 'aevalsrc=0.2*sin(2*PI*1000*t)+0.01|0.2*sin(2*PI*1000*t)-0.02:s=48000:d=5' -c:a pcm_f32le dc-offset.wav
ffmpeg -hide_banner -nostats -i dc-offset.wav -af astats=reset=0 -f null -

Import dc-offset.wav: the column should show 2.000, with Ch 1: +1.000% and Ch 2: -2.000% in the tooltip. FFmpeg should report channel DC offsets close to 0.010000 and -0.020000. Compare individual channel entries, not FFmpeg's separate overall statistic. See the astats documentation.

For a centred control, repeat the generator without +0.01 and -0.02, using another output filename; the column should then show 0.000. In Audacity, inspect the shifted waveform and use DC offset removal on a copy; re-exporting and analysing that copy should produce a value close to zero.