Corpus benchmark

Id Capacity Frames Seconds FrontMs GenMs GenP95 ColdFrontMs ColdGenMs SnrDb

p01 64 56 1.40 5.80 342.40 344.00 22.50 342.00 18.43 p02 96 76 1.90 7.30 585.80 586.90 25.30 585.60 6.06 p03 96 79 1.98 7.50 585.10 586.10 24.60 582.80 22.33 p04 96 92 2.30 7.40 586.70 587.90 24.10 583.80 22.03 p05 128 108 2.70 9.20 1160.60 1161.60 27.00 1125.30 23.44 p06 160 131 3.27 15.50 1369.60 1371.60 31.10 1339.80 24.06 p07 128 101 2.52 9.10 1161.40 1162.70 24.90 1146.40 22.61 p08 128 103 2.58 9.10 1160.60 1162.50 24.30 1149.00 20.60 p09 160 130 3.25 14.80 1370.80 1374.50 31.80 1337.00 23.27 p10 128 125 3.12 9.10 1161.70 1163.20 26.90 1126.80 22.11

Phrases : 9 AudioSeconds : 23.12 GeneratorRtfMean : 0.372 GeneratorRtfP50 : 0.4182 GeneratorRtfP95 : 0.46 GeneratorRtfMax : 0.46 PreparedToAudioMeanMs : 996.7 SnrMeanDb : 22.1 AllPlayed : True

Method

10 phrases (bench/corpus.json, af_heart), each exported at the smallest capacity bucket in {64, 96, 128, 160} that holds its frame count. Per phrase: 1 cold run, then 10 warm repeats of the front and generator contexts, then AudioTrack playback gated on PlaybackHeadPosition. All 9 passing phrases played through the speaker. PreparedToAudioMeanMs is cold front + cold generator, i.e. prepared inputs to first PCM, and excludes the PC-side frontend and har_source.

Notes

• p02 failed the 12 dB SNR gate three times at capacity 96 (6.06 dB) while p03 and p04 pass at the same capacity with 22 dB. This is a per-phrase defect, not a bucket defect, and is excluded from the aggregate. Unresolved.
• Generator time tracks the bucket, not the phrase: 342 / 585 / 1161 / 1370 ms are flat within each bucket (p95 within 2 ms of the mean). RTF is therefore worst for a phrase that just crosses a bucket edge (p05: 108 frames paying 128).