Heartbeat VAD Service

An always-on Android Foreground Service that streams microphone PCM audio through the Silero-based android-vad detector to flag human speech. Tuned for Samsung Galaxy Tab S6 Lite (Android 13/14) with minSdk 26, targetSdk 34, and android:foregroundServiceType="microphone".

Gradle & Build

1. Ensure Android Studio Iguana+ with the Android Gradle Plugin 8.5.x.
2. Open the repository (com.frontieraudio.heartbeat).
3. Sync Gradle, then build & deploy the app module to the device.
4. On first launch grant RECORD_AUDIO and (Android 13+) POST_NOTIFICATIONS.

Runtime Behavior

• HeartbeatService runs as a foreground service (START_STICKY) with no explicit wake locks—the FGS notification is the only keep-alive signal.
• Audio is captured via AudioRecord (16 kHz, mono, PCM_16BIT) using VOICE_RECOGNITION and a single reusable frame buffer.
• Stage 1 VAD (Silero Mode.NORMAL, speech=80 ms, silence=300 ms) gates Stage 2 speaker verification and Stage 3 cloud streaming; only the first speech transition logs VAD: Speech Detected.
• BootCompletedReceiver relaunches the service after device reboots.

Profiling & Validation

Stage 1 success criteria: sustained < 2% CPU with memory flat during the idle listening state. Always capture benchmarks before moving on to Stage 2 (speaker verification) and Stage 3 (cloud streaming).

Android Studio Profiler

1. Deploy the app, grant microphone + notification permissions, and confirm the foreground notification is present.
2. Open Run ▸ Profiler, select the Heartbeat process, then start a Sampled CPU recording (lowest overhead).
3. Allow at least 2 minutes of idle capture to observe steady-state CPU and memory.
4. Note:
   • Average CPU % (target < 1–2%).
   • Java/Kotlin heap + native heap (ONNX Runtime) stability.
5. Export the trace for regression tracking.

Battery Historian

1. Collect a bugreport after an extended idle run (e.g., 8-hour battery test mentioned in the PoC).
2. Load the report into Battery Historian and inspect:
   • Foreground service uptime (should match the test window).
   • Partial wake locks (should be zero since the FGS replaces wakelocks).
   • CPU frequency residency / top consumers.
3. Use these insights to validate that the Stage 1 gate keeps downstream stages off most of the time.

TFLite / ONNX Runtime Benchmarks

Silero runs through ONNX Runtime. To isolate the model cost:

adb shell am force-stop com.frontieraudio.heartbeat
adb shell am start-foreground-service com.frontieraudio.heartbeat/.HeartbeatService
adb shell am profile start com.frontieraudio.heartbeat global

Alternatively, use ONNX Runtime’s ort_benchmark (if flashed onto the device) to replay the Silero .onnx asset under the same CPU affinity and thread count. Target < 2% CPU combined for the service loop plus VAD inference.

Use @benchmark_vad.sh to benchmark as well.

Document each profiling run (date, device build, CPU %, memory MB, Historian observations) in docs/FullAnalysis.md to track regressions before Stage 2 integration.