MansfieldPlumbing
Table of contents

Persistent QNN and AAudio pipeline result

Kokoro-Hexagon 0a03be39Updated 2026-09-24

Historical reference-only experiment. Its QNN contexts, prepared phrase input, and Android host are not the Kokoro-Hexagon product path or a live-speech gate.

Question

Can one Android process keep a QNN front/generator context pair resident while a second in-process PowerShell runspace feeds three chunks through one native AAudio stream, without an unbounded PCM queue or an observed underrun?

Test

The known-good p01 phrase bundle was repeated three times. Repetition is intentional: this isolates process, context, queue, and stream lifetimes from new text-front-end behavior. The main runspace loaded the c64 front and generator contexts once. A second, minimally initialized runspace owned the blocking AAudio writer. A BlockingCollection<float[]> with capacity one was the only inter-runspace PCM queue.

The checked source identities were:

The device startup scripts were copied to a device-local backup before the test runner replaced them. Generated staging files and raw output remained in the adjacent build directory or on the device.

Device result

ContextLoadMs=150.4 Buckets=1
ContextVmHwmKiB=545164
AAudioReady Rate=24000 Channels=1 Format=2 CapacityFrames=768 BurstFrames=48 QueueCapacity=1
FirstAudioQueuedMs=1630.2
Phrase=p01-1 Samples=33600 FrontMs=23.2 GenMs=346.4 EnqueueWaitMs=0.9 SnrDb=18.43 NonFinite=0
Phrase=p01-2 Samples=33600 FrontMs=7.0 GenMs=346.8 EnqueueWaitMs=0.1 SnrDb=18.43 NonFinite=0
Phrase=p01-3 Samples=33600 FrontMs=6.3 GenMs=345.2 EnqueueWaitMs=0.0 SnrDb=18.43 NonFinite=0
AAudioPlaybackStartMs=1713.4 Chunks=3 WrittenFrames=100800 PlaybackFrames=100944 XRunCount=0 PlaybackComplete=True CloseRc=0
FinalVmHwmKiB=590132 ManagedBytes=57638016
PipelineMs=6068.8 Passed=True

The phone speaker played all three chunks. The stream reported 100,800 frames written, at least that many frames presented, zero underruns, complete drain, and a clean close. The three chunks contain 4.2 seconds of audio. The process high-water mark rose by 44,968 KiB after the contexts were resident; this is a whole-process high-water observation, not a per-buffer allocation measurement.

Result and boundary

Pass for this historical diagnostic only: persistent QNN contexts, bounded inter-runspace handoff, synthesis during playback, and one persistent native stream worked for three repetitions of one prepared phrase. This does not validate the PowerShell-authored phoneme-to-PCM product path.

This does not yet prove arbitrary-text phonemization, distinct long-form chunks, cached tensor arenas, seamless subjective boundaries, or a lower peak memory design. Those remain the next gates.