WaveKey

Your voice never leaves the phone.

An Android keyboard that transcribes speech, cleans it up and rewrites it — with two AI models running entirely on the device. Free, with no subscription and no minute counter. No account, no API key, no server, and no network permission in the keyboard process at all.

_{The flow, and the same thing on a real phone.}

See it in motion → nvkudva.github.io/WaveKey

---

What it does

💸 Free, and free of a meter	No subscription, no trial, no minutes to run out of, no tier that unlocks the better model. There is no server to bill you for.
🎙️ Dictate and edit at the same time	The mic stays live while text lands. A pause ends the sentence, not the session — so you can speak, watch it commit, fix a word with your thumb, and keep speaking.
🧠 Two models, both on-device	Parakeet TDT 0.6B turns speech into words. Qwen 3 0.6B turns those words into writing. Neither one leaves the phone.
✨ AI fix, always within reach	One key runs the deterministic rules and then the LLM over what you just wrote — spelling, spacing, casing, clumsy phrasing. Press it again to undo.
⌨️ Continuous fixing as you type	Finished sentences are quietly rescored against the words the decoder ranked second, and the swap is shown to you rather than slipped past you.
🔒 Private by construction	The keyboard process has no INTERNET component. A build-time test fails if one ever appears.
📋 A clipboard that reads like a list	Typed clips — link, image, phone, text — each with a glyph, a size or a host, and how long it has left.

---

Free, and free of a meter

Dictation services charge monthly because they run your voice through their hardware. WaveKey has no hardware to pay for: both models run on your phone, so an hour of dictation costs the battery it takes and nothing else. No account, no trial that expires, no cap.

It is also free in the other sense — GPL-3.0, so every line between your microphone and your text field can be read, audited and forked.

---

The two models

Both are downloaded once, from settings, and then never contacted again.

	Parakeet TDT 0.6B v2	Qwen 3 0.6B
Job	Speech → text	Text → better text
Runtime	sherpa-onnx (ONNX, int8)	LiteRT-LM (mixed int4)
Download	~482 MB, required	~498 MB, optional
Process	the keyboard	a separate :llm process
Licence	CC-BY-4.0	Apache-2.0

Parakeet transcribes each utterance once, when you stop speaking — one accurate pass instead of a stream of guesses that rewrite themselves. Qwen runs behind an AIDL interface in its own process, so a native out-of-memory takes the refiner down and leaves your keyboard standing.

The refiner is optional. Without it you still get dictation and the whole deterministic rules tier; the keyboard tells you plainly when the smart pass could not run rather than pretending it did.

---

Privacy is the architecture

Not a policy — a process split the build enforces.

┌──────────────┐   ┌──────────────┐   ┌──────────────┐
│   keyboard   │   │     :llm     │   │     :ui      │
│  IME + ASR   │   │  Qwen 3 0.6B │   │   settings   │
│              │   │              │   │  + downloads │
│  no network  │   │  no network  │   │   INTERNET   │
└──────────────┘   └──────────────┘   └──────────────┘

ManifestProcessSplitTest fails the build if a component holding INTERNET ever moves out of :ui. Audio is never written to disk, never sent anywhere, and the utterance buffer is cleared when the session ends.

---

Requirements

Android 7.0 (API 24) or newer	The floor the models impose: LiteRT-LM, which runs the refiner, declares minSdkVersion 24. Below it there is no on-device AI to offer, so the keyboard does not install and pretend otherwise.
A 64-bit ARM phone (arm64-v8a)	Every phone since about 2016. The refiner checks for a 64-bit ABI at runtime and stays off without one; the published APK is arm64 only.
A microphone	Dictation adds RECORD_AUDIO to the permissions HeliBoard already asks for.
~1 GB free, ~2 GB with the refiner	482 MB for the speech pack and 498 MB for the LLM, plus room to unpack.
Internet once	For the model download, from the settings process only. The keyboard process has no network component at all.

Nothing else: no account, no API key, no paid service, and no network after the models land.

To build: JDK 21, Android SDK 36. The Gradle 9.6 wrapper and a 4 GB build heap are configured in the repo.

Run it

git clone https://github.com/nvkudva/WaveKey.git
cd WaveKey
./gradlew :app:assembleDebug
# ABI splits produce one APK per architecture; releases ship arm64-v8a
adb install app/build/outputs/apk/debug/WaveKey_1.0-beta-debug-arm64-v8a.apk

"WaveKey" then appears in Android's keyboard list. Enable it, switch to it, and open its settings to download the speech model — the microphone key does nothing until the required pack is installed.

Release signing

There is no runtime configuration. Two environment variables affect release builds only, and only when a keystore exists at ~/.wavekey/release.jks (the old ~/.supervoiceboard/release.jks is still read if that one is absent). Without either file the release build still succeeds and comes out unsigned. With the keystore present, a release build fails unless both passwords are exported — an unsigned APK is uninstallable, so a missing variable stops the build rather than making one.

Variable	Required	What it is
WAVEKEY_STORE_PASSWORD	For release builds	Keystore password for release signing
WAVEKEY_KEY_PASSWORD	For release builds	Key password for the signing alias
WAVEKEY_KEY_ALIAS	No	Signing alias; defaults to wavekey

---

How it is built

WaveKey is a fork of HeliBoard 4.1 (base commit 9f5bb63). Typing — layouts, glide typing, dictionaries, themes — is HeliBoard's, unchanged. This repo adds the voice and AI layer.

Module	What lives there
core/	Pure Kotlin JVM, no Android. The decisions: DictationStateMachine, TranscriptCleaner, TextFixer, PackInstaller. About two thirds of the module by line count is tests.
voice/	The Android half. VoiceSessionController runs the state machine's effects against AudioCapture and the Parakeet recognizer. It references no HeliBoard class, so it could be mounted in another IME.
llm/	Qwen 3 under LiteRT-LM in the :llm process, behind ILlmRefiner.aidl.
app/	The binding. VoiceController is the only class that writes to the InputConnection; AiFixKey mounts the AI fix key.

Issues about the voice layer belong on this repo's tracker; issues about typing belong upstream.

Status

1.0-beta. Working today: dictation, on-device cleanup, the AI fix key, sentence rescoring, and model download and install from settings.

Verified on the current tree:

Unit tests	core 932 · voice 72 · app 214 — all passing
Android lint	0 errors
Release build	assembles and signs
Model integrity	both packs pinned to an immutable revision and a SHA-256 the installer checks

CI runs the core, voice and app unit tests, a debug assemble, Android lint, and an emulator UI QA suite.

Known gaps

Honest ones, from a code review of voice/, llm/ and the voice code in app/:

• Late refinement can delete characters typed after a commit. replaceUtterance removes previous.length characters before the cursor without checking that they are still the text it committed.
• llm/ has no unit tests, and voice/ — which holds most of the fork's concurrency — is only part-way there: the root rules, the level meter, the silence timeout and idle release are covered; VoiceSessionController is not.
• A timed-out bindService leaves the binding in place, pinning the :llm process and its model until something else tears it down.
• No coroutine exception handler on the IME-process scopes, so an uncaught throw in a finalize, refine or fix coroutine reaches the default handler and takes the keyboard with it.

Nothing about speed, accuracy or battery has been measured, and dictation is English only.

License

GPL-3.0-only — see LICENSE. The AOSP Keyboard base is Apache-2.0 (LICENSE-Apache-2.0), the launcher icon is CC-BY-SA-4.0 (LICENSE-CC-BY-SA-4.0), and the default icon set is MIT (LICENSE-MIT-fluent-icons).