Polished Recognition

A voice keyboard powered by your own AI.

Tap the mic, speak, and polished text appears — transcribed by the speech provider you choose, refined by the language model you choose, with a prompt you write. Bring your own API key (or run a local model) and take full control of your voice input.

Polished Recognition installs as a voice keyboard: switch to it, and it starts listening immediately — no extra tap, no learning a new typing UI.

Gboard will never use this. Gboard's microphone is hardcoded to Google's voice typing — it cannot be redirected to any third-party voice input, including Polished Recognition. To use Polished, switch to a keyboard that delegates its mic and select Polished Recognition as its voice input: the AOSP keyboard (LineageOS and derivatives), HeliBoard (F-Droid only), Fossify Keyboard, or OpenBoard.

The voice keyboard — recording	Your pipeline, your prompts
Your providers, your models	Dictate, send, polished text appears

How it works

1. Switch to the Polished keyboard → recording starts instantly
2. Speak naturally — pause/resume anytime mid-sentence; need to type something? Switch to your normal keyboard — the dictation waits and continues when you're back
3. Tap send → audio goes to your STT provider (Whisper on GROQ, OpenAI, …)
4. The LLM polishes the text using your custom prompt — fix filler words and punctuation, restructure, or translate to another language
5. Flawless text lands in the app you're typing in

Don't need the polish? Uncheck LLM in the keyboard bar: the LLM is skipped entirely and the plain transcription is inserted — zero extra latency.

Why it's different

• BYO provider — or use a free one. 19 presets (OpenAI, GROQ, OpenRouter, Cortecs, Google AI, DeepSeek, xAI, Mistral, local Ollama/LM Studio, …). GROQ's free tier gives you Whisper STT plus a strong LLM — a complete STT+polish pipeline at zero cost.
• Not a walled garden. No account, no registration, no central server. Your audio goes only to the provider you configured — or nowhere at all, if you run Ollama locally.
• Custom prompts. The system prompt is fully editable, with variables for source and target language. Make the LLM format markdown, translate to French, or just fix punctuation.
• Translation built in. Pick a target language and dictation arrives translated. Languages are freely editable — long-press to add your own.
• Pause & resume. Get interrupted mid-dictation? Pause, handle it, resume — the entire recording buffers, nothing is lost. You can even switch to your typing keyboard and back: the dictation survives the round trip.
• Works everywhere. Any text field, any app — no integration needed. It also registers as the system voice-input service for keyboards that delegate their mic button (HeliBoard, Fossify/OpenBoard, the AOSP keyboard) — note that Gboard never delegates: its mic is hardcoded to Google's own voice typing.
• Searchable model pickers. Type to filter hundreds of models by substring, with per-provider caching. No infinite dropdown scrolling.
• Clean output. LLM "reasoning/thinking" blocks are stripped automatically — only the finished text gets inserted.

Installation

See the Installation Guide for step-by-step setup: Play Store installation, configuring providers, enabling the voice keyboard, and device-specific notes.

Become an Alpha Tester

Polished Recognition is in closed testing on Google Play — and I'm still looking for testers! Two steps, same Google account for both:

1. Join the testers group: https://groups.google.com/g/polished-recognition-alpha
2. Accept testing on the opt-in page → install from the Play Store

Prefer F-Droid?

Stay opted in — Google's 14-day production-eligibility clock only runs while at least 12 testers stay continuously opted in. Details in the Installation Guide.

Don't be shy with feedback — report bugs and ideas as GitHub Issues, or open a Pull Request if you've fixed something. I happily accept contributions.

Providers

Configured via presets or custom URLs — anything speaking the OpenAI API contract works:

Provider	Type	Base URL
GROQ Whisper	STT	https://api.groq.com/openai/v1/
OpenAI Whisper	STT	https://api.openai.com/v1/
OpenAI	LLM	https://api.openai.com/v1/
OpenRouter	LLM	https://openrouter.ai/api/v1/
Google AI Studio	LLM	https://generativelanguage.googleapis.com/v1beta/openai/
GROQ	LLM	https://api.groq.com/openai/v1/
Cortecs	LLM	https://api.cortecs.ai/v1/
DeepSeek	LLM	https://api.deepseek.com/v1/
xAI	LLM	https://api.x.ai/v1/
Mistral	LLM	https://api.mistral.ai/v1/
Together AI	LLM	https://api.together.xyz/v1/
DeepInfra	LLM	https://api.deepinfra.com/v1/
Fireworks	LLM	https://api.fireworks.ai/inference/v1/
Cerebras	LLM	https://api.cerebras.ai/v1/
Perplexity	LLM	https://api.perplexity.ai/
HuggingFace	LLM	https://router.huggingface.co/v1/
NVIDIA	LLM	https://integrate.api.nvidia.com/v1/
Ollama (local)	LLM	http://localhost:11434/v1/
LM Studio (local)	LLM	http://localhost:1234/v1/

Or add any custom provider by entering a base URL + API token. Models are fetched dynamically from each provider's /v1/models endpoint (with a fallback to free-text entry for providers that don't support it).

Features

Voice keyboard (IME)

• Instant recording — switching to the keyboard starts the mic right away
• Two-line control surface — settings gear, target language, LLM toggle on top; keyboard switch (the X), delete-word, discard, pause and send below
• One-tap keyboard switch — the X in the bottom row pauses the dictation and hands the field straight back to the keyboard you came from; if no typing keyboard is available it opens the keyboard settings so you can enable one. It always switches — the "switch back after send" setting only controls the automatic switch after a successful send
• Discard is its own button — the trash icon clears the audio and timer without leaving the keyboard, so throwing a take away never costs you the way back
• Dictation survives keyboard switches — while you type on the other keyboard the recording is preserved (paused); switching back auto-resumes and appends to the same audio
• Live stage display — shows exactly what's happening: Recording, Transcribing (STT), Polishing (LLM)
• Recording pulse — the send button breathes with your voice
• Works in any app — any text field, any app, no integration needed

Your pipeline, your rules

• LLM switch — checked (default) the LLM polishes every dictation: with a target language it translates, LLM, keep language polishes in the language you spoke; uncheck the LLM toggle in the keyboard bar to bypass the LLM completely and insert the plain transcription. The toggle lives in the keyboard bar (not in Settings) and remembers its state
• Editable prompts — System Prompt with {{source_language_clause}} / {{target_language_clause}} variables, plus the translation clause with {{target_language}}; reset individually or all at once
• Custom target languages — the language dropdown accepts any language name; long-press to edit or add, and your languages are remembered
• Token & model validation — test tokens with a minimal chat request, fetch model lists on demand, cached per provider
• Request & response logging — rotating JSON logs of the STT and LLM exchanges on-device, pullable via adb for debugging

Privacy

No account, no analytics, no tracking, no proprietary SDKs. The only network traffic is your own API calls to the endpoints you configured. Everything sensitive — keys, prompts, logs — stays on your device. See the Privacy Policy.

Troubleshooting

• Recording fails / no audio? Check that RECORD_AUDIO is granted — the app requests it on first use.
• Model dropdown empty? Enter a valid token and press Validate & Fetch Models; model lists are fetched from the provider (some providers only allow free-text model entry).
• Voice-input service greyed out? Some keyboards/vendor builds restrict the system voice service — use the Polished keyboard directly instead.
• Gboard's mic doesn't launch it? That's by design — Gboard hardcodes Google's voice typing and never uses external voice keyboards. Switch to the AOSP keyboard (LineageOS), HeliBoard (F-Droid), Fossify, or OpenBoard. See the Installation Guide.
• More help? See the Installation Guide.

For developers

Command	Result
./gradlew assembleRelease	Build release APK (minified, release-key signed when a local keystore is present)
./gradlew installRelease	Build + install release APK via ADB
./gradlew test	Run all unit tests

Stack: Kotlin, Retrofit + OkHttp, AudioRecord (16 kHz mono → in-memory WAV), plain XML settings UI on Theme.DeviceDefault.DayNight. Min SDK 30. No DI framework, no Room, no Compose — deliberately small and fast.

Technical details (architecture, decisions, conventions, pitfalls) are maintained in docs/ai/ — start with index.md, then HANDOFF.md.

License

MIT