level op• データ種: signal → signal
• 呼び出し: import acoustics; acoustics.apply_weighting(x, rate, kind='A') (または opsacoustics.get("apply_weighting"))
信号に A / C / Z 周波数重み付けをゼロ位相で掛ける。
> 以下の詳細説明は原文のままです —— 要約と見出しは訳出済み。
The weighting is applied as a real, even gain in the frequency domain, so it
introduces no phase distortion and no group delay — the result is aligned
sample-for-sample with the input, which a recursive filter implementation
would not be.
Measured: a 1 kHz sine at 16 kHz (16000 samples, exactly 1000 periods) is
returned unchanged by both A and C weighting — max absolute difference
1.078e-13 for A and 1.225e-13 for C — because both curves are exactly 0 dB
at 1 kHz by construction. A 100 Hz sine of amplitude 1.0 comes back with
amplitude 0.110373 under A weighting, against the closed form
`10**(-19.1428/20) = 0.110373`.
`kind="Z"` returns a copy, unchanged.
**A tone that is not a whole number of periods in the record reads too
loud, by up to 17 dB, and nothing raises.** The multiplication is over the
record's own DFT, which treats it as periodic; a tone that does not close
on itself leaks across every bin. That leakage would be harmless if the
weighting were flat, but A weighting spans about 40 dB between 20 Hz and
1 kHz, so a sidelobe 40 dB below a 31.5 Hz tone arrives at 1 kHz weighted
40 dB *higher* and takes over the sum. Measured, 0.5 s at 48 kHz, error
against the closed-form `A(f)` for a pure tone:
========== ============= ========== ==============
f (Hz) periods error (dB) bin-centred?
========== ============= ========== ==============
22.0 11.0 +0.0000 yes
31.5 15.75 +7.7986 no
20.5 10.25 +17.2116 no (worst, 20-200 Hz)
63.0 31.5 +0.1121 no
100.0 50.0 +0.0000 yes
1000.0 500.0 -0.0000 yes
========== ============= ========== ==============
31.5 Hz is a nominal one-third-octave centre, so this is a path a real
measurement walks into rather than a contrived one. The error is always
*positive* — leakage only ever adds power at frequencies the curve favours.
Two things confirm the diagnosis is dynamic range and not arithmetic. The
same 31.5 Hz tone under C weighting, whose tilt over the same span is a
few dB rather than forty, is off by only +0.0493 dB. And lengthening the
record to where the tone *does* close on itself removes it entirely: at
31.5 Hz the error is +7.7524 dB over 0.25 s, +7.7986 over 0.5 s, +0.4615
over 1 s, and -0.0000 over 2 s and 4 s (63 and 126 whole periods).
Two candidate cures were measured (error in dB against the closed form,
0.5 s at 48 kHz):
=================================== ======== ======== ========
treatment 31.5 Hz 20.5 Hz 22.0 Hz
=================================== ======== ======== ========
as implemented (rectangular) +7.7986 +17.2116 +0.0000
zero-pad x4 (linear convolution) +5.5620 +14.3352 +0.7969
Hann window, corrected for its gain +0.0534 +0.1841 +0.1505
=================================== ======== ======== ========
Padding barely helps — zero-padding a tone puts an abrupt edge into the
record and an edge is broadband. A Hann window does essentially cure it,
turning +17 dB into +0.18 dB, at the cost of the bin-centred columns which
go from exactly 0 to about 0.15 dB. So why is it not the default?
Because it would trade a loud error for a quiet one. `L_eq` is an
*energy average over the record*, and a window is not energy-preserving for
anything that is not stationary. Measured with Z weighting (so the window is
the only thing acting) on a 50 ms 1 kHz burst inside a 0.5 s record, all
three placements being `-13.0103` dB unwindowed as they must be:
============== ============ ===========
burst position Hann (dB) difference
============== ============ ===========
start -36.0587 -23.05
centre -8.8218 +4.19
end -36.0587 -23.05
============== ============ ===========
A window makes the answer depend on *where in the record the sound happened*,
which is precisely the "plausible wrong number" this module refuses to ship
by default. So the rectangular behaviour stays, and the Hann estimate is
available by asking for it: `equivalent_level(..., window="hann")`. Use it
when the record is stationary and tonal — which is exactly when the leakage
bites — and never when the level of a transient is the point.
A cure with neither cost is a different implementation entirely: the
standard cascade of A-weighting biquads in the time domain, which would give
up the exact-0-dB-at-1-kHz-by-construction property this function is built
on, and the zero group delay promised above.
Also worth doing: give the analysis enough record that the content is
many periods long, prefer durations that are whole multiples of the period
you care about, and read a low-frequency A-weighted level from
:func:octave_spectrum (which reports per-band power, so leakage is visible
as energy in bands where none belongs) rather than from a single number.
Raises `ValueError: everything :func:_as_signal` refuses, an unknown
`kind, rate <= 0`.
• acoustic_condition_monitoring ファミリ ガイド
• サンプルデータ カタログ(DL URL / ライセンス) — 2-D は skimage.data(BSD/public)+ 合成、3-D は実データ源(Stanford/PDS 等)の DL URL。
• 演算子の来歴・参考文献 — この op 族の元になった研究/手法の出典。
• アルゴリズムの正典(著者・年)と用途は上記ファミリ使い方ガイドに記載。
• acoustic_condition_monitoring — py -3.11 examples/acoustic_condition_monitoring.py
signal を入力に取れる)stft · envelope_spectrum · spectral_kurtosis · cepstrum · angular_resample · order_spectrum · octave_spectrum · weighting_response
level)octave_bands · octave_spectrum · weighting_response · equivalent_level · percentile_level
*Provenance: acoustics.py — ACOUSTICS operator registry. この per-op ノートは tools/opdocs.py md が自動生成(手編集しない)。*
© 2026 Kazufumi Furuse — Fullseye operator documentation. Licensed under Apache-2.0.