marmalade stt
Open-source speech-to-text for Android

Speak, and your phone types. Your voice never leaves it.

Private, offline voice typing for your keyboard's mic key and for any app that asks Android for speech input. Your words appear while you're still talking.

Free, open source, no ads, no accounts. Android 9 or later. Coming soon to F-Droid.

Voice: Kokoro. Text: Parakeet Ultra's real output.

Works where you type

Marmalade STT plugs into Android's own voice input, so you keep the keyboard and the apps you already use.

Your keyboard's mic keyTurn on Marmalade voice typing, and the mic key opens a small Marmalade panel that types what you say.
The speak-now popupApps with a "speak now" button get a small Marmalade sheet that returns your words when you finish.
Android's speech recognizerPick Marmalade STT as the speech recognition service, and apps that use it hear you through Marmalade.

Tested with

  • FUTO Keyboard
  • HeliBoard
  • FlorisBoard
  • AnySoftKeyboard
  • Unexpected Keyboard
  • Android's own keyboard
  • Fossify Keyboard
  • OpenBoard
  • Thumb-Key

Gboard's mic key always starts Gboard's own voice typing, but Android's keyboard switcher still reaches Marmalade. The setup guide shows how for each keyboard you use, and the full table is in the README.

Support development

Marmalade STT is free, with no ads and no accounts, and the app is the same whether you sponsor it or not. If it is useful to you, you can sponsor its development. Donations are voluntary and not tax-deductible.

Sponsor on GitHub

What you get

Everything happens on your phone. The only download is the speech engine you choose.

Built on open speech models

Marmalade runs open models directly on your phone. You download only the engine you choose, once, from GitHub, and every file is checked against a pinned SHA-256 hash.

The engines run on sherpa-onnx (Apache-2.0, built from source) and ONNX Runtime (MIT). Full credits and license texts are in the app and in NOTICE.md.

100 languages, the right engine for each

Pick the language you speak, or leave it on Auto to follow your phone, and the setup guide recommends the engine that suits it.

Engine picked for your language
Hello
English
Parakeet Ultra
  • 25 European languagesParakeet UltraEnglish, Bulgarian, Croatian, Czech, Danish, Dutch, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Russian, Slovak, Slovenian, Spanish, Swedish, Ukrainian
  • 日本語Parakeet JapaneseJapanese
  • Hindi, Arabic, VietnameseNemotron 3.5
  • Chinese, KoreanQwen3-ASR
  • Every other languageWhisper Large v3 TurboWith more delay

On phones short of memory or speed, the guide suggests a lighter engine: Parakeet 110M for English, or Canary 180M Flash for Spanish, German and French.

The app itself is in 21 languages: English, Arabic, Chinese (Simplified and Traditional), Czech, Dutch, French, German, Hindi, Indonesian, Italian, Japanese, Korean, Polish, Portuguese (Brazil), Russian, Spanish, Swedish, Turkish, Ukrainian and Vietnamese.

Screenshots

A setup guide, an engine for your phone, and a panel that types what you say.

Welcome screen of the setup guide: private by design
Private by design
The Marmalade panel typing into a messaging app
Typing from the mic key
Engines tab with the engine recommended for this phone
An engine for your phone
Engines tab in dark mode
Engines, dark
Setup guide step: dictate from your keyboard
Setup for your keyboards
End of speech and text cleanup settings
When it stops, how it writes
Setup guide finished: microphone, speech model, voice keyboard and speech recognition service ready
All set
Dictate tab with dictated text
The Dictate screen

Private by design

No analytics, no tracking, no crash reporting, no accounts, and no server of ours. Speech is recognized on the phone, and audio is never saved or uploaded. The internet is used only to download the speech engines you choose, from GitHub, and every file is checked against a pinned SHA-256 hash. Android backup is turned off, so nothing is copied to the cloud.

Privacy policy

Open source

Marmalade STT is MIT licensed, and the app you install contains no GPL or LGPL code. Its speech runtime, sherpa-onnx, is built from source with the text-to-speech parts left out. Every component's license is listed in the app under Settings › About › Open-source licenses.

Source code on GitHub · Issues · NOTICE.md