Canonical fp16 baseline — frozen 2026-09-22
Historical QNN reference measurement, not a product model, build, or speech gate.
Device: Galaxy S23 (SM8550, Hexagon V73), QAIRT 2.46.0.260424, AndroidSMA .preview host. Phrase: af_heart, 131 frames (3.27 s audio), 160-frame capacity. Burst performance vote. Method: contexts loaded and tensors bound once, one warm-up execute, 20 timed executes per stage.
| Artifact | SHA-256 |
|---|---|
| front_c160_fp16.bin | 9D445E0BFF7D0087… (full hash in known-good.json) |
| gen_nin_gb_c160_fp16_vtcm8.bin | EFAC4EA2598843E1… |
Job=kokoro-speak
FrontMs=36.6 GenMs=1351.6 (first call)
NonFinite=0 AudioSnrDb=24.06
BenchFront N=20 MeanMs=14.9 P50Ms=14.5 P95Ms=15.3 MinMs=14.2 MaxMs=19.8
BenchGen N=20 MeanMs=1372.9 P50Ms=1373.0 P95Ms=1375.1 MinMs=1370.0 MaxMs=1378.0
Passed=True
Generator RTF (warm mean): 1.3729 / 3.275 = 0.419.
Comparison candidate (plain graph, 8 MB VTCM, gen_banal_c160_fp16_vtcm8.bin FD867EF1E7A8A138…): BenchGen N=20 MeanMs=1562.6 P50Ms=1562.4 P95Ms=1564.9; AudioSnrDb=23.99.
Reproduction with playback-head gate (repository driver, 2026-09-22)
front sha=9D445E0BFF7D0087 gen sha=EFAC4EA2598843E1
FrontMs=32.5 GenMs=1346.8
NonFinite=0 AudioSnrDb=24.06
PlayState=Playing WrittenFrames=78600 PlaybackFrames=78600 PlaybackComplete=True
BenchFront N=20 MeanMs=14.7 P50Ms=14.5 P95Ms=15.3 MinMs=14.0 MaxMs=16.3
BenchGen N=20 MeanMs=1372.4 P50Ms=1372.5 P95Ms=1375.5 MinMs=1368.2 MaxMs=1377.4
Passed=True