Whisper

Whisper is an independent Android speech-recognition app built around the open-source Whisper model. It can run as a small input method, register as Android’s system voice-recognition service, answer RecognizerIntent requests from compatible apps, or work as a standalone recorder that transcribes speech and can translate supported languages into English.

This page covers F-Droid’s 3.8 build. We matched the 20.6 MB APK to F-Droid’s signed index and inspected the fixed source, model downloader and recognition paths. We did not time inference or compare accuracy across phones, accents and languages; those depend heavily on the chosen model and device CPU.

Model download and offline boundary

The installer contains vocabulary/filter data but not the recognition models. On first launch, the app requires all three current files from the developer’s Hugging Face repository: whisper-tiny.en.tflite (41,486,616 bytes), whisper-base.TOP_WORLD.tflite (107,564,368 bytes) and whisper-small.TOP_WORLD.tflite (307,408,944 bytes). Together they are 456,459,928 bytes, about 435 MiB. F-Droid therefore flags the package NonFreeNet, even though the model files themselves are published under MIT terms.

Once the files are present, recording and inference run locally. The source loads the selected TFLite file from app storage and feeds it recorded audio; it does not need to upload each dictation to a speech API. At every startup the downloader checks each file against a hard-coded MD5 value and asks to download the set again if a required file is missing or does not match. Network permission remains for that acquisition. “Offline voice typing” therefore describes use after setup, not installation on a permanently disconnected device.

The app can use the tiny English-only file, the faster multilingual base file or the slower multilingual small file, but version 3.8 downloads all three rather than letting setup fetch only one. The default is the multilingual whisper-small.TOP_WORLD model with automatic language selection. Hardware makes a large difference: a slower phone may take noticeable time after the 30-second maximum recording ends.

IME, voice service and translation

As an IME, Whisper presents a record control, a translate-to-English toggle and inserts the result into the current field. A keyboard such as HeliBoard can invoke it from its microphone button. Because it becomes an input method, Android shows the standard warning that keyboards can observe typed fields; this project’s code performs recognition locally, but users should still treat any enabled IME as sensitive software.

It can also be selected as Android’s voice-input service. Some manufacturers show only their hard-coded Google or Samsung choices. Upstream documents an ADB command that sets Whisper’s recognition service when the settings UI hides it; that changes a secure Android setting and is not required on devices that list the app normally.

Standalone mode can translate supported speech into English. That is Whisper’s speech-translation task, not a general text translator between arbitrary language pairs. Each recording is limited to 30 seconds. Bluetooth microphone access depends on Android version, permission and the connected audio route.

Permissions and file facts

The manifest declares microphone and audio-settings access, Bluetooth, internet, notifications, boot, vibration, a general foreground-service permission and the microphone foreground-service type, plus package visibility. Our static scan found no known tracker SDKs. No tracker finding is not an accuracy or security certification; it only describes recognized packaged libraries.

About this file

• This is F-Droid’s source build of 3.8, signed with F-Droid’s key.
• It needs Android 9.0 or later and contains 32- and 64-bit ARM native code; there is no x86 build.
• Its measured SHA-256 is f97d8087dc7a81a1175a08f8858d808815b3554f031ffaa982a6fd04ac861e7c, matching F-Droid’s signed index.
• The app is MIT-licensed. It is an independent Whisper-based project, not an official OpenAI Android application.

Flagged by F-Droid:

• NonFreeNet: Relies on or promotes a proprietary network service.

CPU types in this build: arm64-v8a (64-bit ARM, almost every phone since 2017); armeabi-v7a (32-bit ARM, older and budget phones).

Versions currently kept in the F-Droid repository. We host and check only the newest one.