Table of contents

Corpus float playback

Kokoro-Hexagon 0a03be39Updated 2026-09-24

Historical QNN reference playback only; this does not validate the direct PowerShell phoneme-to-PCM product path.

Date: 2026-09-24 Target: physical Samsung Galaxy S23, SM8550, Hexagon V73 Model: Kokoro-82M Weight path: FP16 QNN AOT contexts Voice: af_heart Workload: frozen ten-phrase corpus, ten warm executions per stage

The corpus ran after moving quality comparison and diagnostic WAV creation out of the first-audio path and writing float PCM directly to the Android audio sink. All ten phrases synthesized and played completely. Nine passed the 12 dB reference-quality gate; p02 remained the known quality-gate failure at 6.06 dB and was excluded from passing aggregates.

Metric Earlier diagnostic path Float playback path
Passing phrases 9 9
Passing audio 23.12 s 23.12 s
Generator RTF mean 0.3725 0.3724
Generator RTF p50 0.4147 0.4180
Generator RTF p95 0.4611 0.4601
Prepared input to playback start, mean 2,565.5 ms 1,648.2 ms
Mean audio SNR 22.10 dB 22.10 dB
Complete playback yes yes

Mean prepared-input-to-playback handoff decreased by 917.3 ms, or 35.8%, without a material generator or quality change. The shortest 64-frame phrase decreased from 1,650.6 ms to 772.1 ms. Longer phrases still pay for complete waveform generation and copying before playback; persistent chunked synthesis and native AAudio are required to remove that structural delay.

Per-phrase results

ID Capacity Audio Generator mean Playback start SNR Gate
p01 64 1.400 s 343.0 ms 772.1 ms 18.43 dB pass
p02 96 1.900 s 587.3 ms 1,159.4 ms 6.06 dB quality
p03 96 1.975 s 587.7 ms 1,202.5 ms 22.33 dB pass
p04 96 2.300 s 589.3 ms 1,159.3 ms 22.03 dB pass
p05 128 2.700 s 1,161.5 ms 1,844.0 ms 23.44 dB pass
p06 160 3.275 s 1,369.0 ms 2,163.6 ms 24.06 dB pass
p07 128 2.525 s 1,161.7 ms 1,855.5 ms 22.61 dB pass
p08 128 2.575 s 1,162.2 ms 1,833.3 ms 20.60 dB pass
p09 160 3.250 s 1,369.3 ms 2,166.0 ms 23.27 dB pass
p10 128 3.125 s 1,163.2 ms 1,837.7 ms 22.11 dB pass