VoiceNotes PoC — Android

Test pipeline: mic → Gemma 4 E2B → transcript + tags + intent

Setup

1. Download the model (Gemma 4 E2B INT4, ~1.5GB):

   • From HuggingFace: google/gemma-4-e2b-it (MediaPipe format)
   • Push to device: adb push gemma4-e2b-it-int4.bin /data/data/com.voicenotes.poc/files/

2. Build & install:

   ./gradlew installDebug

3. Run: Hold the record button → speak → release → wait for Gemma output

What it measures

• Cold model load time
• Inference latency (ms) per recording
• Quality of transcript + tag extraction in Italian/English
• Intent detection accuracy

Files

• AudioRecorder.kt — 16kHz PCM capture + WAV conversion
• GemmaEngine.kt — MediaPipe LlmInference wrapper + JSON parser
• MainActivity.kt — hold-to-record UI, shows raw output

Note on audio modality

MediaPipe GenAI Tasks multimodal audio API support depends on the SDK version. If audio input isn't available yet, the engine falls back to text-only inference (describe the audio path in the prompt). Check MediaPipe release notes for tasks-genai >= 0.10.22.