Persistent AAudio stream reuse
Historical QNN-backed stream-lifecycle test, not a product speech gate.
Date: 2026-09-24
Target: physical Samsung Galaxy S23, SM8550
Input chunk: p01, 33,600 frames, 1.40 seconds
Pipeline: Kokoro-82M, FP16 QNN AOT contexts, af_heart
One PowerShell runspace opened one native AAudio stream and wrote the same validated PCM buffer twice as two sequential chunks. The stream was started only for the first write and drained only after both writes.
| Metric | Result |
|---|---|
| AAudio open | 82.4 ms |
| Prepared input to first playback start | 905.7 ms |
| Stream format | 24 kHz mono float PCM |
| Buffer capacity / burst | 768 / 48 frames |
| Chunks | 2 |
| Frames written | 67,200 |
| Frames consumed at drain | 67,440 |
| Underruns | 0 |
| Playback complete | yes |
| Audio SNR | 18.43 dB |
| Close result | 0 |
The second chunk reused the live stream and paid no additional open or start operation. This result proves the native audio queue substrate only; the speech waveform was synthesized once. The next long-form gate is producing successive chunks through cached QNN contexts in the same runspace.