First light

Historical QNN/QAIRT decoder playback only. This is not PowerShell-authored phoneme-to-PCM Kokoro synthesis, not the model-less appliance, and not a product speech gate. The prepared acoustic inputs and QNN contexts must not become release dependencies.

Device: Samsung Galaxy S23 (SM-S911U, SM8550, Hexagon V73), QAIRT 2.46.0.260424. Host app: AndroidSMA .preview, arm64 Debug, built from AndroidSMA a320c45 with <uses-native-library android:name="libcdsprpc.so" android:required="false"/>.

Phrase: af_heart, 131 frames (3.27 s), 160-frame capacity.

First vocalization (PowerShell iSTFT)

Job=kokoro-speak
FrontMs=68.1 GenMs=4688.7
TotalMs=14482.0 Samples=78600 Seconds=3.27
NonFinite=0 AudioSnrDb=24.25
PlayState=Playing Played=True
Passed=True

Whole decoder on HTP (iSTFT in graph, burst vote, 8 MB VTCM)

Job=kokoro-speak
Perf=burst SetRc=0
FrontMs=38.8 GenMs=1567.0
TotalMs=2345.0 Samples=78600 Seconds=3.27
NonFinite=0 AudioSnrDb=23.99
Passed=True

AudioSnrDb is measured against the full-length (unmasked) PyTorch decoder. The fp32 CPU run of the same masked design scores 24.2 dB, so HTP adds no measurable error beyond the design.