WhisperInput

Offline voice input panel & keyboard with punctuation for Android, experimental, powered by Whisper AI & Kõnele components.

Voice input is supported in English. Multilingual input can be used, see in Installation.

Features

• Works as a voice keyboard (input method editor), a voice input panel, or an assistant app.
• On-device speech recognition, offline.
• Auto-start, auto-stop, audio cue option.

Usage tips

• To set the app as a web search assistant (long press Home button to open voice input), open the app -> Settings gear icon -> Recognition services (system UI). A system menu should open for selecting the assistant app, for example in Samsung UI it's Device assistance app. Select Whisper Input in the list.
• You can switch your keyboard to Whisper Input voice keyboard in the system keyboard list. Click the keyboard icon when done to switch back.

Installation

Requirements: Java, Android SDK.

Initial setup:

• Put your keystore.jks to the project's root folder for signing the app.
• Create a signing.properties in the project's root folder with keystore.jks credentials:

signingStoreLocation=../keystore.jks
signingStorePassword=<keystore.jks password>
signingKeyAlias=<keystore.jks alias>
signingKeyPassword=<keystore.jks key password>

• (Optional) To replace the included English-only speech model with a bigger or multilingual one, replace ggml-tiny.en.bin in the assets/models folder with another .bin type model from the whisper.cpp model list. The models without .en. are multilingual. Tiny or base size models recommended. Note: a multilingual model is expected to perform worse in English than the .en. model of the similar size.

Run:

git clone https://github.com/alex-vt/WhisperInput.git
cd WhisperInput/
./gradlew assembleRelease

Install app/build/outputs/apk/release/app-release.apk on Android device.

Development

Points of interest:

• OfflineRecognitionService.kt
• WhisperRecognitionModel.kt

Possible issues:

• Permissions handling. The permissing model from the Kõnele components may be falling back behind new Android versions requirements, that may result in failure to record voice. A workaround is to allow permissions manually in Android's app settings.
• The voice model may take a few seconds to load - the first recording may be inconsistent. Optimally the ready status of the model at WhisperRecognitionModel.kt:35 should be reflected in the recording UI.

Underlying projects

• Kõnele, a voice inputs integration app for Android, by Kaljurand, Apache-2.0 license
• speechutils, a voice recognition library for Android, by Kaljurand, Apache-2.0 license
• whisper.cpp, a performance-tuned build of a speech model, by ggerganov, MIT license
• whisper, a speech model, by OpenAI, MIT license

License

MIT license