Offline voice typing for Android, with a choice of speech models, recoverable recordings, personal dictionary corrections and optional local transcript cleanup.
This personal fork keeps FUTO Voice Input's voice keyboard and speech-recognition activity. Orukeet is the default recognizer. Models download separately; recognition and cleanup then run on your phone.
Download the APK · What differs from FUTO · Model guide · Release notes · Report a problem
Stable 1.4.6 promotes the tested Share beta 7 code, with clearer Model Options, Parakeet Redux and the Nemotron download fix.
- Install the signed APK from GitHub Releases. You need Android 8.0 or newer and an ARM64 device.
- Open settings, grant microphone permission and follow the voice-input setup.
- Open Model Options, keep Orukeet or choose a model, and download its files.
- Select Voice Input Moonshine through your keyboard's voice-input control or Android's input-method picker, then try a short dictation.
The app uses org.futo.voiceinput.moonshine, separate from upstream FUTO Voice
Input. Install updates over the existing app to retain settings, models and
history. Do not uninstall or clear app storage.
Use Check for updates in settings or Obtainium for stable updates. Installation, updates and keyboard setup covers the details. This app is distributed through GitHub Releases and is not listed in F-Droid's main repository.
This fork keeps FUTO's Android voice-input flow and adds:
- More recognition choices, with Orukeet as the default and streaming options for text while you speak.
- Recording history to copy transcripts and retry saved audio.
- Personal dictionary corrections, bulk paste and text-file imports.
- Optional local English cleanup with S1-mini by Superwhisper and Harper.
- A microphone waveform, an optional bottom-positioned popup and local diagnostics.
Upstream also develops multiple recognition models and streaming. The dated upstream comparison explains which versions were reviewed. Model choice and results depend on your language and phone; we have no controlled benchmark establishing a universal speed or accuracy advantage.
Start with Orukeet. If you want text while speaking, choose a streaming model. Try the same recording with different models before downloading several large ones.
| Your use | Starting option |
|---|---|
| Everyday dictation in a supported European language | Orukeet |
| Live English text | Moonshine Small |
| Live multilingual text | Nemotron 3.5 Multilingual |
| Arabic dictation | Cohere Transcribe with Arabic selected |
| Original FUTO recognition path | Legacy Whisper |
Model Options shows languages, download size, unpacked model size and readiness. The screenshot was captured on a Samsung Galaxy S25 Ultra for beta 7; the interface is unchanged in stable 1.4.6.
The full model guide includes all models, source links, storage requirements and the other two screenshots. Download size and installed model size are different from RAM use.
Speech recognition and cleanup run locally after downloading their models. Diagnostics stay local until you export and share a report yourself.
Recording backups are on by default, with 24-hour retention. They include canceled and failed attempts. You can turn backups off, change retention or delete recordings in settings. Uninstalling or clearing app data removes them. Audio history and recovery explains the controls.
Personal Dictionary applies explicit text corrections. S1-mini provides optional English rewriting after recognition and is off by default. Turn it off for non-English dictation on models that cannot report a language. Dictionary and cleanup instructions cover imports, language handling and Harper settings.
The maintainer uses a Samsung Galaxy S25 Ultra. Testing on that phone does not establish compatibility with every phone, model, keyboard or headset.
- The dedicated FUTO Keyboard provider setup entry is hidden in 1.4.6 pending upstream support. See keyboard setup.
- Redux Android inference and comparative model benchmarks still need device validation.
- The Poco no-speech report remains unresolved.
For a bug, include the app version, phone/Android version, model, keyboard or editor, reproduction steps and what happened. Export a diagnostic report if useful, and review it before sharing. Do not attach private audio or transcripts unintentionally.
The release notes record verification and remaining limits. GitHub Issues tracks bugs and planned work.
I like FUTO Voice Input's interface and wanted to keep it while experimenting with other local speech models. Orukeet and Cohere have worked well in my everyday use. I test changes on my phone and welcome feature ideas and device reports.
The code changes for this fork are written by AI agents, directed by me. I choose the work, test the app and provide feedback. AI tools are used in development; dictation uses the local models described above. Project background and AI disclosure explains the workflow and preserves upstream and third-party attribution.
- User guide: installation, keyboards, history, dictionary, cleanup, popup and diagnostics.
- Model guide: model catalog, downloads and screenshots.
- Development guide: builds, contributions and releases.
- Fork background: upstream comparison, project history, AI development and attribution.
- Phone testing procedure: collect comparable measurements on your device.
This fork preserves the FUTO Source First license and notices. It is not affiliated with or endorsed by FUTO. Model weights and runtimes have their own terms; see attribution and licenses.