Repository navigation
Effects Limiter Comparison 2026 11 05
48 kHz・stereo・256 framesで、準備2秒、測定4秒分の音声を処理した。各試行は実時間10ms以上かつ全区間1周以上のburn-in後に状態をリセットし、再び準備して測定する。39条件をブロックごとに無作為化し、10ブロック、ブロック間5秒、CPU0固定で実施した。seedは2026110501。入力・全出力配列を先に確保し、処理区間にI/Oや確保を入れない。
性能試験のLimiter設定はceiling−1dB、Input Gain+6dB、release100ms、左右max共通gain。音響試験では入力波形の振幅を直接指定するためInput Gain0dBを使う。
これは同じ数値設定による異なる音響方式の比較である。減衰量、過渡応答、歪みが違うため、速度だけで方式を置き換える判断はしない。
logs/assets/effects-2026-11-05/dynamics/results/performance-figures/limiter-cpu.png
| 方式 | Bass | Pad | Drums |
|---|---|---|---|
| instant 0 ms | 2.787 | 2.793 | 2.806 |
| box 1 ms | 7.929 | 8.333 | 7.974 |
| triangle 1 ms | 9.002 | 9.435 | 9.139 |
| compensated 1 ms | 10.309 | 9.834 | 10.350 |
| box 5 ms | 7.940 | 8.439 | 7.973 |
| triangle 5 ms | 9.069 | 9.556 | 9.132 |
| compensated 5 ms | 10.101 | 9.468 | 10.294 |
単位はns/stereo frame。bypassは bass 0.453, pad 0.466, drums 0.451 ns/frameで、上の値から差し引いていない。図のerror barは各条件の平均の95%区間、方式間の判定は次のブロック内ペア差による。
39条件全体で事前登録した42比較にHolm補正を適用した。差は候補−参照、比率は候補平均/参照平均。
| 参照 → 候補 | 平均差 ns/frame | 差の95%区間 | t (df=9) | Holm p | 比率 |
|---|---|---|---|---|---|
| limit.instant.max.0.bass → limit.box.max.1.bass | +5.142 | [+5.089, +5.195] | 218.983 | 1.63e-16 | 2.845 |
| limit.instant.max.0.bass → limit.box.max.5.bass | +5.153 | [+5.080, +5.227] | 158.753 | 2.14e-15 | 2.849 |
| limit.instant.max.0.bass → limit.triangle.max.1.bass | +6.216 | [+6.182, +6.250] | 411.406 | 6.33e-19 | 3.231 |
| limit.instant.max.0.bass → limit.triangle.max.5.bass | +6.283 | [+6.200, +6.365] | 172.312 | 1.16e-15 | 3.255 |
| limit.instant.max.0.bass → limit.compensated.max.1.bass | +7.523 | [+7.454, +7.591] | 248.214 | 5.41e-17 | 3.700 |
| limit.instant.max.0.bass → limit.compensated.max.5.bass | +7.314 | [+7.265, +7.363] | 336.968 | 3.73e-18 | 3.625 |
| limit.box.max.1.bass → limit.triangle.max.1.bass | +1.074 | [+1.007, +1.140] | 36.630 | 6.26e-10 | 1.135 |
| limit.box.max.1.bass → limit.compensated.max.1.bass | +2.381 | [+2.307, +2.455] | 72.724 | 1.96e-12 | 1.300 |
| limit.box.max.5.bass → limit.triangle.max.5.bass | +1.129 | [+1.050, +1.209] | 32.119 | 1.62e-09 | 1.142 |
| limit.box.max.5.bass → limit.compensated.max.5.bass | +2.161 | [+2.072, +2.250] | 54.910 | 2e-11 | 1.272 |
| limit.instant.max.0.pad → limit.box.max.1.pad | +5.541 | [+5.451, +5.630] | 139.906 | 6.43e-15 | 2.984 |
| limit.instant.max.0.pad → limit.box.max.5.pad | +5.646 | [+5.580, +5.712] | 192.113 | 4.85e-16 | 3.022 |
| limit.instant.max.0.pad → limit.triangle.max.1.pad | +6.642 | [+6.563, +6.722] | 189.487 | 5.33e-16 | 3.379 |
| limit.instant.max.0.pad → limit.triangle.max.5.pad | +6.763 | [+6.715, +6.811] | 320.346 | 5.73e-18 | 3.422 |
| limit.instant.max.0.pad → limit.compensated.max.1.pad | +7.041 | [+6.894, +7.188] | 108.296 | 5.95e-14 | 3.521 |
| limit.instant.max.0.pad → limit.compensated.max.5.pad | +6.675 | [+6.619, +6.731] | 269.100 | 2.68e-17 | 3.390 |
| limit.box.max.1.pad → limit.triangle.max.1.pad | +1.102 | [+0.975, +1.228] | 19.642 | 6.39e-08 | 1.132 |
| limit.box.max.1.pad → limit.compensated.max.1.pad | +1.500 | [+1.315, +1.686] | 18.323 | 7.85e-08 | 1.180 |
| limit.box.max.5.pad → limit.triangle.max.5.pad | +1.117 | [+1.019, +1.216] | 25.647 | 9.03e-09 | 1.132 |
| limit.box.max.5.pad → limit.compensated.max.5.pad | +1.029 | [+0.924, +1.134] | 22.189 | 2.9e-08 | 1.122 |
| limit.instant.max.0.drums → limit.box.max.1.drums | +5.169 | [+5.101, +5.236] | 172.614 | 1.16e-15 | 2.842 |
| limit.instant.max.0.drums → limit.box.max.5.drums | +5.167 | [+5.112, +5.222] | 212.457 | 2.08e-16 | 2.842 |
| limit.instant.max.0.drums → limit.triangle.max.1.drums | +6.333 | [+6.244, +6.422] | 161.241 | 1.93e-15 | 3.257 |
| limit.instant.max.0.drums → limit.triangle.max.5.drums | +6.327 | [+6.247, +6.407] | 178.750 | 8.74e-16 | 3.255 |
| limit.instant.max.0.drums → limit.compensated.max.1.drums | +7.544 | [+7.462, +7.626] | 207.983 | 2.45e-16 | 3.689 |
| limit.instant.max.0.drums → limit.compensated.max.5.drums | +7.488 | [+7.388, +7.588] | 169.323 | 1.29e-15 | 3.669 |
| limit.box.max.1.drums → limit.triangle.max.1.drums | +1.165 | [+1.070, +1.259] | 27.923 | 5.18e-09 | 1.146 |
| limit.box.max.1.drums → limit.compensated.max.1.drums | +2.376 | [+2.330, +2.421] | 117.751 | 2.92e-14 | 1.298 |
| limit.box.max.5.drums → limit.triangle.max.5.drums | +1.159 | [+1.064, +1.254] | 27.675 | 5.18e-09 | 1.145 |
| limit.box.max.5.drums → limit.compensated.max.5.drums | +2.320 | [+2.227, +2.414] | 55.924 | 1.79e-11 | 1.291 |
10ブロックは分布の形を詳しく断定するには少なく、ペア差の正規性・lag-1相関も全JSONへ記録した。cooldownは負荷状態を離す操作であり独立性の証明ではない。有意差がない比較を同等とは扱わない。
即時版は軽く遅延0、先読み版は数倍の計算量で制御を先行させる。1msから5msへの変更は状態容量を増やす。最大値queueと移動平均の区間平均の償却計算量は1フレーム当たりO(1)だが、1サンプルでは最大O(D)のqueue更新があり得る。ここでDは先読みサンプル数であり、CPU表の区間平均を1サンプルの最悪時間保証とは扱わない。Box・Triangle・補償IIRのCPU差は、異なるゲイン曲線の費用として読む。
全試行・実行順・環境・対応t検定、固定plan、共通測定規約、共通runner、Rust CLI、処理器、図・表生成。バイナリSHA-256は a3811f93e6cedbfb14e98ce28e68672f709fb637146af6903c4a6e53e73d12f0。
入力Gainを統一したTrue-peak sidechain、逆振幅領域の平滑化、数値型・出力再利用と10分の精度へ、検出精度・回復曲線・数値安定性の比較をまとめた。全f32の移動和をそのまま使う場合と、和だけf64にする場合を区別する。
3秒burn-inによる別の確認測定では、初回の固定方式・条件・実行ファイルを用い、各族の効果量と検定を独立に保存する。