Native AAudio speaker path

Kokoro-Hexagon 0a03be39Updated 2026-09-24

Historical QNN-backed audio-sink test. The native AAudio result does not make the QNN model path a product dependency or prove direct-path speech.

Date: 2026-09-24 Target: physical Samsung Galaxy S23, SM8550 Phrase: p01, 33,600 frames, 1.40 seconds Pipeline: Kokoro-82M, FP16 QNN AOT contexts, af_heart

PowerShell bound the Android NDK AAudio C ABI directly through libaaudio.so. The stream used blocking float PCM writes and no managed Android.Media.AudioTrack object.

Metric	Result
Front	41.2 ms
Generator	341.3 ms
Whole waveform ready	745.5 ms
AAudio stream open	81.6 ms
Prepared input to playback start	910.2 ms
Sample rate	24,000 Hz
Channels	1
Format	float PCM (`AAUDIO_FORMAT_PCM_FLOAT`)
Buffer capacity	768 frames
Frames per burst	48
Frames written / consumed	33,600 / 33,600
Underruns	0
Audio SNR	18.43 dB
Close result	0

Playback completed through the phone speaker and the result gate passed. The stream was opened for this single diagnostic phrase, so this result does not claim a first-audio improvement over the earlier managed static-buffer test. The verified AAudio stream is intended to stay open in the persistent long-form runspace, where stream-open and first-start costs are paid once.

ABI source: AOSP frameworks/av commit 9e7dd63dfff0cc967f025ea9e27a299aaa99fd69, AAudio.h blob 25ad5f8ef976c33b8670e4217d3f984c13096829.