Voice Inbox

Android application for incrementally transcribing a folder of audio recordings into an existing plain-text (.txt) or Markdown (.md) file.

Workflow

1. Download and verify the speech model on first launch.
2. Select an existing writable .txt or .md file.
3. Select a folder containing audio recordings.
4. Review discovered recordings in the New tab and choose Transcribe all.
5. The application processes new recordings sequentially in foreground work and appends one entry per successful file containing its filename, recording time, and recognized text.

File selection remains disabled until the model is installed and loaded. The application preserves existing text and separates each new entry with one blank line. Folder scanning runs when the application starts and when Refresh is selected. It scans direct children only and does not traverse nested folders.

Incremental Catalog

The application stores folder entries and processing state in a private SQLite database. The New tab contains new, changed, processing, and failed files. The Processed tab contains successful files. A failed file is excluded from Transcribe all until its individual Retry action is selected.

Files are identified by provider URI and considered changed when their reported size or modification time changes. Providers that do not expose reliable metadata can therefore hide a content-only change. Full content hashing is not performed. Missing files remain in private catalog history but are hidden from both tabs.

Every successful reprocessing appends a new entry. The application never searches, replaces, or deletes previous Markdown content because the document may be edited independently. A process failure after an append succeeds but before SQLite records success can make the same file eligible again and produce a duplicate entry; generic document providers cannot make those two updates atomic.

Supported Audio

The application uses Android platform decoders and adds no third-party media decoder. The tested input set is:

• 8-bit and 16-bit linear PCM WAV
• AAC-LC and HE-AAC in M4A/MP4
• MP3
• ADTS AAC
• mono or stereo FLAC
• Ogg Vorbis
• Ogg Opus
• AMR-NB and AMR-WB

Decoded audio is downmixed and resampled to 16 kHz mono before recognition. Encrypted, DRM-protected, malformed, and unsupported multichannel inputs are rejected.

Model And Device Requirements

• ARM64 Android device
• Android 7.0 (API 24) or later
• About 704 MiB free for first model installation
• Network access for the first model download

The pinned int8 Parakeet model download is about 640 MiB. After verification it is stored in the application's private no-backup directory. Recognition runs locally after installation.

Build

Requirements:

• Android SDK and NDK 30.0.14904198
• Rust toolchain
• cargo-ndk

./gradlew assembleDebug
./gradlew testDebugUnitTest

The Gradle build extracts ONNX Runtime, builds the Rust JNI library for arm64-v8a, and packages the required native libraries.