WhisperVault — Private, Verified Voice Recognition

MainActivity · RecognizeActivity (manual) · RecognizeActivity (auto, ∞ visible) · Settings · IME manual · IME auto · IME continuous (∞ active)

WhisperVault is a security-hardened voice recognition IME built on OpenAI Whisper. It runs entirely offline — your audio never leaves the device.

Security properties

• 100% offline inference — no network access during recognition; audio stays on-device
• Model integrity verification — SHA-256 hashes checked at extraction and at every load across all five entry points; mismatches are surfaced to the user before any inference runs
• Zip Slip protection — canonical path validation on every entry before writing during model extraction
• Minimal permissions — QUERY_ALL_PACKAGES removed; package visibility declared via <queries> only
• Hardened manifest — setup and settings activities are not exported; visibleToInstantApps removed from recognition activity
• Release build hardening — R8 minification enabled with ProGuard rules that preserve ONNX Runtime reflection; ProGuard also strips diagnostic logging
• No diagnostic leakage — transcription output and target package names are gated behind BuildConfig.DEBUG; not present in release builds
• Thread-safety — recognitionCancelled and isVerifying declared volatile; RecordBuffer uses ConcurrentLinkedQueue for continuous mode and AtomicReference with consume-once getAndSet(null) semantics for single-buffer mode
• 47 tests — 35 JVM unit tests (including Robolectric manifest checks) and 12 Espresso/UIAutomator E2E tests run against a real emulator with real ONNX inference

WhisperVault is a hardened fork of whisperIMEplus. It has three distinct entry points, each with a different UI:

Entry point	How it launches	UI	Screenshots
Standalone app (MainActivity)	Tap the app icon	Full-screen: text output area, copy button, Append/Translate toggles, mic button	01
Voice input overlay (WhisperRecognizeActivity)	Another app requests speech via RecognizerIntent (e.g. a search bar, assistant, or keyboard's mic key)	Floating dialog over the caller: cancel (✕), mode buttons (A / ∞), mic. No keyboard controls — the caller handles the returned text. Closes automatically after transcription.	02, 03
IME keyboard (WhisperInputMethodService)	User selects WhisperVault as their keyboard, or taps the mic button in HeliBoard	Bottom keyboard panel: mode strip (A / ∞ / ⌨ / EN), status line, mic button. Android draws the standard delete and enter keys below. Stays open while focus is in a text field.	05, 06, 07

The voice input overlay (02, 03) intentionally has no keyboard controls because it is not a keyboard — it captures speech and returns the text string to the calling app via Android's RecognizerIntent API. The standalone app and IME both write text directly.

As a standalone app it can also translate any supported language to English.

Initial Setup

Upon launching WhisperVault for the first time, you will need to download the Whisper model from Hugging Face and install it. Voice recognition works entirely offline, ensuring your privacy and convenience.

Please note that for use as voice input (not as IME) there is a separate settings activity which can be accessed from Android settings (System > Languages > Speech > Voice Input). There you can activate the app as voice input and then click the settings button.

If after installation you do not find WhisperVault as voice input or only see a limited list (hard-coded ones like Google/Samsung)

• enable USB debugging
• type adb shell settings put secure voice_recognition_service io.github.nick_tgcs.whispervault/com.whisperonnx.WhisperRecognitionService

Using WhisperVault

To get the most out of WhisperVault, follow these simple tips:

• Press and hold the mic button while speaking (manual mode), or use auto/continuous mode
• Pause briefly before starting to speak
• Speak clearly, loudly, and at a moderate pace
• In continuous mode, just keep talking — pauses between sentences are transcribed automatically

Recording Modes

WhisperVault supports three recording modes, controlled by the A (Auto) and ∞ (Continuous) buttons:

Mode	How to activate	VAD behavior	Stops when	Mic button color
Manual	Press & hold mic	Filters silence (pauses don't consume 30s buffer)	User releases button	Blue → Amber → Red → Light blue
Auto	Tap A	Stops on silence (one-shot)	Silence detected	Blue → Amber → Red → Light blue
Continuous	Tap A then ∞	Hands off utterance on pause, keeps mic open	User taps A again, or privacy stop	Blue → Amber → Red (stays red during processing)

stateDiagram-v2
    [*] --> Manual
    Manual --> Auto : Tap A
    Auto --> Continuous : Tap A or Tap ∞
    Continuous --> Manual : Tap A
    Auto --> Manual : Tap A
    Continuous --> Auto : Tap ∞
    Auto --> Manual : Privacy stop
    Continuous --> Manual : Privacy stop

Loading

Button behaviors:

Button	Tap when A is off	Tap when A is on (auto)	Tap when A is on (continuous)
A	Activates auto mode	Switches to continuous mode	Deactivates — back to manual
∞	Not visible	Switches to continuous mode	Switches back to auto (one-shot)
Mic	Press & hold for manual recording	Disabled (A controls recording)	Disabled (∞ controls recording)

Mic Button Visual States

The mic button changes color to show what's happening:

State	Color	Meaning
Idle	Blue	Not recording — waiting for user action
Listening	Amber	Mic is open, VAD hasn't detected speech yet
Recording	Red	VAD detected speech, actively capturing
Processing	Light blue (auto/manual) or Red (continuous)	Transcribing audio

Auto Mode Flow

flowchart TD
    A["User taps A"] --> B["A button ON, ∞ button appears"]
    B --> C["Mic: AMBER, Status: 'Listening…'"]
    C --> D["Recorder starts (STOP_ON_SILENCE)"]
    D --> E{"VAD detects speech?"}
    E -->|Yes| F["Mic: RED, Status: 'Recording…'"]
    F --> G{"VAD detects silence?"}
    G -->|Yes| H["Mic: LIGHT BLUE, Status: 'Processing…'"]
    H --> I["Utterance handed off for transcription"]
    I --> J["Recording stops"]
    J --> K["Text committed to input"]
    K --> L["Auto mode ends, reset to Manual"]
    L --> M["Mic: BLUE (idle), Status hidden"]

    style A fill:#1976D2,color:#fff
    style F fill:#CF6679,color:#fff
    style H fill:#64B5F6,color:#000
    style C fill:#FFB74D,color:#fff
    style M fill:#1976D2,color:#fff

Loading

Continuous Mode Flow

flowchart TD
    A["User taps ∞"] --> B["∞ button ON"]
    B --> C["Mic: AMBER, Status: 'Listening…'"]
    C --> D["Recorder starts (CONTINUOUS)"]
    D --> E{"VAD detects speech?"}
    E -->|Yes| F["Mic: RED, Status: 'Recording…'"]
    F --> G{"VAD detects pause?"}
    G -->|Yes| H["Mic stays RED, Status: 'Processing…'"]
    H --> I["Utterance handed off for transcription"]
    I --> J["Recorder resets, starts new utterance"]
    J --> K["Status: 'Listening…'"]
    K --> E

    I --> L["Whisper transcribes"]
    L --> M["Text committed to input"]
    M --> N["Loop continues"]

    D -.->|Privacy stop| Z["Mic OFF, reset to MANUAL"]
    E -.->|Privacy stop| Z
    F -.->|Privacy stop| Z
    K -.->|Privacy stop| Z

    style A fill:#1976D2,color:#fff
    style F fill:#CF6679,color:#fff
    style C fill:#FFB74D,color:#fff
    style K fill:#FFB74D,color:#fff
    style H fill:#CF6679,color:#fff
    style Z fill:#D32F2F,color:#fff

Loading

Privacy: Microphone Stops on Loss of Focus

Continuous mode keeps the microphone open, which makes privacy safeguards critical. The microphone stops immediately when the user is no longer actively interacting with a text field:

flowchart LR
    A["Continuous recording active"] --> B{"User interaction ends?"}
    B -->|Phone locks| C["onFinishInputView"]
    B -->|App switch| D["onStartInput TYPE_NULL"]
    B -->|IME hidden| E["onWindowHidden"]
    B -->|IME destroyed| F["onDestroy"]
    B -->|Input closed| G["InputConnection = null"]
    B -->|User taps ⌨| H["switchToPreviousInputMethod"]

    C --> I["stopAllRecording"]
    D --> I
    E --> I
    F --> I
    G --> I
    H --> I

    I --> J["Mic: BLUE, Mode: MANUAL, Status: hidden"]

    style I fill:#D32F2F,color:#fff
    style J fill:#1976D2,color:#fff

Loading

No recording ever continues in the background. There is no scenario where the microphone is active while the user is not looking at a text field with WhisperVault as their active keyboard.

You can also let the app translate your input to English using the translate button.

License

This work is licensed under GPLv3. Original code © woheller69. Modifications © 2026 nick-tgcs.

• This app is a fork of whisperIMEplus, licensed under GPLv3
• whisperIMEplus is based on whisperIME, published under MIT license
• It uses code and the Whisper ONNX models from RTranslator
• It uses code from Whisper-Android project, published under MIT license
• It uses OpenAI Whisper published under MIT license. Details on Whisper are found here.
• It uses Android VAD, which is published under MIT license
• It uses Opencc4j, for Chinese conversions, published under Apache-2.0 license
• At first start you need to download the Whisper model from HuggingFace, which is published under MIT license