stft_cola_check — ACOUSTICS transform op

• データ種: なし → table(引数だけで決まる op —— 画像やデータの入力を取らない)

• 呼び出し: import acoustics; acoustics.stft_cola_check(window='hann', win=256, hop=None) (または opsacoustics.get("stft_cola_check"))

使い方

その(窓, ホップ)の組は COLA を満たすか、どれだけ厳密に満たすか。

> 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。

COLA — the analysis windows summing to a constant over the hop lattice — is

what lets plain overlap-add work without a division. It is *not* required by

:func:istft, which is weighted, but it is required by anything that

overlap-adds modified frames without renormalising, and getting it wrong

produces a periodic amplitude ripple at `rate/hop` Hz that looks like

tremolo rather than like a bug.

Returns a dict: `cola (bool), constant` (the mean of the overlap sum),

`max_deviation (absolute), relative_deviation, nola` (bool),

`min_squared_sum`, plus the geometry.

Measured (periodic windows, `relative_deviation` of the plain sum):

======== ==== ==== ================== ======== ====

window win hop relative_deviation constant COLA

======== ==== ==== ================== ======== ====

hann 256 128 4.44e-16 1.00 yes

hann 256 64 2.22e-16 2.00 yes

hann 256 85 1.48e-03 1.506 no

hamming 256 128 2.06e-16 1.08 yes

blackman 256 128 1.91e-01 0.84 no

blackman 256 64 3.97e-16 1.68 yes

boxcar 256 128 0.00e+00 2.00 yes

======== ==== ==== ================== ======== ====

The two blackman rows are the useful ones: the same window is COLA at

hop = win/4 and 19 % off at hop = win/2, so "which window" is not the

question — the pair is. The boxcar row was worth measuring rather than

assuming: a rectangular window at 50 % overlap sums to exactly 2 and is

COLA, which is the opposite of the usual intuition about it. Note also that

the constant is not 1 in general — an overlap-add that does not divide by it

is off by a *gain*, which is the failure that looks like a working system.

Raises `ValueError: unknown / all-zero window, hop` outside

`[1, win], win outside [2, MAX_WINDOW]`.

詳しい使い方ガイド

• acoustic_condition_monitoring ファミリ ガイド

参考(サンプルデータ・文献)

• 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。

実行できる例(この op を実際に呼ぶ検証済みサンプル)

• acoustic_condition_monitoring — py -3.11 examples/acoustic_condition_monitoring.py

型が繋がる次の op(table を入力に取れる)

istft

同カテゴリ(transform)

stft · istft

*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.