Offline Speech to Text (OSTT)

app.offlinespeechtotext
by Offline Speech to Text

Private, fully offline speech to text for Android

Offline Speech to Text turns speech into text on your Android phone. After downloading or importing a speech model, recognition works without a network connection or account. Audio is processed on the device. Use the built-in voice keyboard to dictate into text fields. Keyboard dictation can keep recording when the keyboard hides or the screen turns off, with an ongoing notification and Stop control. If its original app is unavailable, the result stays in private recovery. Background continuation requires enabled notifications; recording limits still apply. You can also select the app as your speech recognition provider in keyboards and apps that support Android's provider interface. Some keyboards use their own service and do not offer this choice. Share an audio file or supported video note with the app to transcribe its audio. Copy, share or save the transcript. Files decode in blocks to limit audio-buffer memory. Other apps can request file transcription through documented Android contracts. Live subtitles can transcribe audio from other apps when Android allows audio capture. On Android 12 and later, optional Bluetooth microphone support can use a connected headset or LE Audio hearing aid. The recording screen identifies the active microphone and lets you switch. If the Bluetooth input stops, capture can continue on the phone microphone with a cue and status message. Use separate dictionary views for preferred words and explicit heard-to-written corrections. Choose the pause before automatic stopping. Transcript recovery helps preserve text when an editor changes or closes. Password fields are protected from automatic insertion. The app download does not bundle a speech model. Home guides microphone access and model setup. Speech models shows installed models first, or a recommendation based on your language, storage and phone when none are installed. Tap Download and install to fetch a model. Other models, local import and model preferences are available from the menu. The download picker offers eleven tested choices, including Parakeet v3 Q8 and Canary 1B v2. Supported multilingual models can use a Languages I speak list to limit detection to your selected languages. Downloaded models remain in private app storage through ordinary updates. Compatible transcribe.cpp GGUF models can also be imported from local storage. The app is free and open source under the MIT License, with a native Android interface, a Rust core and the transcribe.cpp speech engine. Detailed troubleshooting is off by default. Enable it for one hour, 24 hours or until manually disabled. Logs stay private and bounded; exporting them requires your action.

First release: Sep 11, 2026, 33 total releases.

Most recent release: Oct 30, 2026.

Website Repo

Appears in 2 app stacks.

2,100 sats / 1 zaps received in the past year.

Sats Received

Underlying data available via MCP: app_zaps, app_releases.

Zap Count

Underlying data available via MCP: app_zaps, app_releases.

Releases

  • Oct 30, 2026 0.23.0
    Keep keyboard dictation recording when the keyboard hides, with a foreground notification and Stop control. Preserve text privately when its original app is unavailable. Prefer the microphone of the headset selected for media, and allow more time for LE Audio microphone routing to become ready.
  • Oct 29, 2026 0.22.1
    Report a busy file-transcription queue correctly and allow cancelled jobs to release their slots for later requests. Skip file jobs cancelled before startup. Harden file-model worker creation and ensure extractor cleanup even when an audio codec throws during teardown.
  • Oct 29, 2026 0.22.0
    Add optional Bluetooth microphone support with a phone-microphone fallback. Transcribe shared audio and video notes, share or save transcripts, and let other apps request file transcription. Keep caller-supplied audio separate from microphone dictation. Improve request completion, cancellation and resource cleanup.
  • Oct 25, 2026 0.21.0
    Add Canary 1B v2 to model downloads and a Languages I speak setting for supported multilingual models. Preserve your preferred language for short clips. Improve English filler removal while protecting non-English words and personal corrections. Add Donate in Settings to open a Lightning wallet using this app’s address, with an explicit copy fallback. Reduce audio copying during long-recording preparation.
  • Oct 21, 2026 0.20.0
    Warm amber controls on black in the dark theme. Installed speech models appear before download recommendations. A simpler model screen keeps import and preferences in the menu. The download picker offers ten tested models, including Parakeet v3 Q8. Download progress shows the model name. Removed duplicate setup text and fixed Automatic language warnings.
  • Oct 21, 2026 0.19.0
    Simpler setup for microphone permission and an offline speech model. Show one model recommendation first, with installed models and advanced choices available separately. Rework the dictionary into preferred words and explicit corrections, with protected drafts and clearer recovery. Use a compact true-black dark interface. Keep recording, subtitles and advanced controls in Settings.
  • Oct 20, 2026 0.18.1
    Fix dictionary corrections and preferred phrases failing to match uppercase Greek text. Preserve quoted text, paths and existing Unicode lowercase expansion handling.
  • Oct 20, 2026 0.18.0
    New microphone artwork and accessible keyboard status announcements. More reliable model downloads with pinned sources. Long dictation no longer gains an unnecessary decode window from final silence padding. Dictionary entries with letters and digits now restore their exact spelling. Explicit corrections and preferred terms survive filler cleanup. English cleanup removes “um” as well as “uh”, while preserving capitals, quotes, code and paths. Select English for cleanup with multilingual Parakeet.
  • Oct 18, 2026 0.17.0
    Detailed troubleshooting now offers Enable for 1 hour, Enable for 24 hours, Enable forever, then Cancel in a fixed order. Forever keeps detailed logging enabled until you switch it off; disabling basic logging also revokes it. The warning, private metadata-only collection and bounded log remain. Updated privacy documentation and verified consent, expiry and cancellation behavior.
  • Oct 18, 2026 0.16.0
    Updated the speech engine to transcribe.cpp 0.4.0. Audio files now decode in bounded blocks, with overlapping transcription windows and recoverable partial results. Dictation, file transcription and subtitles share one model, with priority for waiting voice input. Improved speech detection, silence trimming, keyboard switching, accessibility and Android 16 compatibility. Reduced native debug data in the APK. Existing dictionaries, transcript recovery and downloaded models remain supported.
  • Oct 6, 2026 0.14.0
    Choose one hour or 24 hours of content-free detailed troubleshooting, with automatic expiry and bounded retention. Diagnostic timestamps now reflect event time and distinguish recognition from callback delivery. Reuse the microphone-meter animation and stop it when its view detaches. Speech recognition, models, and privacy defaults are preserved.
  • Oct 6, 2026 0.13.1
    Maintenance update: simplify content-free diagnostic validation and marker writes. Speech recognition, models, settings and privacy defaults remain unchanged. Includes safe dictation preparation and troubleshooting controls from 0.13.0.
  • Oct 5, 2026 0.13.0
    Add optional early processing of long voice-keyboard recordings using existing overlapping windows, with one final result and no brief-pause cuts. Short dictation stays unchanged. Add separately warned, one-hour detailed troubleshooting with private timing, CPU and thermal metadata, never audio or transcripts. Improve no-text recognition messages and stale callback checks.
  • Oct 4, 2026 0.12.2
    Fix caller-audio Stop draining and report input failures instead of returning truncated recognition. Add private, opt-in outcome diagnostics. Recognition models and settings are unchanged.
  • Oct 1, 2026 0.12.1
    Correct diagnostic labels for recording and cancellation, record content-free recognition error codes, and identify the build at diagnostic session startup. Includes the previously prepared dictionary cleanup. Speech recognition behavior is unchanged.
  • Sep 30, 2026 0.12.0
    Personal dictionary: exact names such as Wi-Fi and O’Reilly, safer quote/email/code protection, one consistent list per transcription, and safer concurrent saves. Choose a 2, 3 or 5 second trailing pause when Auto-stop is enabled. Existing defaults, offline processing and installed models are preserved.
  • Sep 24, 2026 0.11.0
    English dictation now removes a standalone “uh” after recognition while preserving nearby words, sentence capitalization, and quoted speech. Other languages and filler-only results stay unchanged. For multilingual Parakeet, choose English under the model language setting to enable this English-only cleanup; Auto mode cannot reliably identify the language. Models, settings, and transcripts remain on your device.
  • Sep 24, 2026 0.10.1
    Voice input now waits through Android's device-wide microphone unblock prompt instead of treating the blocked interval as a pause in speech. Fixed automatic stopping in the voice keyboard, improved the first-install message when no speech model is present, and made preferred-word spelling cleanup available across supported model families. Ordinary words are not automatically replaced with similar-sounding names.
  • Sep 23, 2026 0.10.0
    Preferred phrases now work with Parakeet and other non-Whisper models: exact matches use your chosen spelling and capitalization, including names such as White Noise. The preferred-word list also handles close versions of Nostr and a few common spoken forms of npub. This stays private and on-device. Larger or ambiguous recognition errors still need an optional exact correction; OSTT does not rewrite text based on broad sound-alike guesses.
  • Sep 23, 2026 0.9.0
    Add a simpler personal dictionary: enter preferred words once, without listing every misrecognition. Whisper uses bounded recognition hints; Parakeet and other models conservatively correct very close spellings of single words after transcription. Exact correction pairs remain available for phrases and larger errors. Make the dictionary screen denser and clarify what each model can do. Use warm, high-contrast confirmation colors in light and dark themes.
  • Sep 22, 2026 0.8.0
    Keep long dictation smooth across natural pauses. OSTT now preserves each ordinary recording as one continuous utterance and lets the speech runtime maintain its own long-audio context, preventing short hesitations from creating false sentence boundaries. Very long input to other offline model families uses overlapping, reconciled windows. Diagnostic logging remains off by default and content-free. When enabled, Settings can save the bounded log as a text file, share a temporary read-only snapshot through Android's Sharesheet, or clear it immediately.
  • Sep 22, 2026 0.7.0
    Faster long dictation with model-aware background processing while you speak. Long recordings now finish sooner without dropping audio: OSTT keeps the original recording until the final result and automatically retries the complete message if background processing cannot finish. Rapid app or editor changes also release an old recording immediately, avoiding false microphone errors when the next transcription starts.
  • Sep 21, 2026 0.6.3
    Keep completed voice-keyboard transcriptions in the active composer when Compose or another reactive UI recreates its input connection. OSTT now retries brief connection replacement and inserts into Android's currently focused safe composer instead of opening its recovery screen. If no writable composer exists, the exact transcript remains privately recoverable without interrupting the current app.
  • Sep 21, 2026 0.6.2
    Fix automatic voice-keyboard recording when switching between apps or text fields. An editor transition before the first audio buffer no longer produces a false microphone-permission error. When Android keeps the voice keyboard visible for a new editor, automatic recording starts there correctly, while any real captured result still uses private recovery if its original field is gone.
  • Sep 21, 2026 0.6.1
    Fix a startup crash in compact installs that do not yet have a speech model. The home screen now waits to initialize transcription until a downloaded, imported, or retained model is actually available, while the native loader handles a missing packaged model as a normal setup state instead of allowing an Android file exception to escape.
  • Sep 21, 2026 0.6.0
    Upgrade the offline transcription engine to transcribe.cpp 0.2.3 for current decoder fixes and more efficient resource handling. German-primary phones now recommend the German-tuned Parakeet primeLine model, while the multilingual Parakeet model remains the balanced default elsewhere. Manual model imports are more resilient: copying, verification, and atomic installation now continue safely across screen rotation, with clearer progress and failure cleanup.
  • Sep 20, 2026 0.5.0
    Never lose a finished voice-keyboard transcription when Android refreshes or closes the original text field. Same-editor connection refreshes now insert normally; otherwise OSTT saves one private recovery result with copy, keep, and delete controls. The APK is now small: first setup recommends and downloads a verified model once, and ordinary app updates keep it. Models are grouped and explained by family, with the largest comfortable option for this phone highlighted. Download guidance uses phone resources and system language, with an App Info shortcut for optional Network removal afterward.
  • Sep 20, 2026 0.4.1
    Keep final words by giving offline decoders a short inference-only silence tail after recording stops. Use simple original/replacement fields for exact personal-dictionary corrections, with no rule syntax to learn. Render every voice-keyboard icon in white for a consistent high-contrast layout. Keep optional diagnostic logging responsive with bounded append-and-compact storage.
  • Sep 19, 2026 0.4.0
    Add a private personal dictionary with Whisper hints and safe exact corrections for every model. Show an amber progress ring while transcription is processing, improve optimized ARM backend detection, and document custom voice UI integration. Keep keyboard switching and live subtitles safe on older Android versions.
  • Sep 18, 2026 0.3.1
    Model selection is now the first section on the Speech models screen. Recommendations and verified downloads follow underneath, with manual import and source links kept as secondary options. Normal ready-state status and retry controls stay hidden until they are needed.
  • Sep 17, 2026 0.3.0
    Offline-first model setup and a faster, denser voice workflow. - Includes the recommended Parakeet model so transcription works without a download. - Adds guided model recommendations, verified in-app downloads, and practical device-fit estimates. - Keeps download progress and cancellation on the model card that started the download. - Adds compact pause, cancel, insert, and insert-and-switch keyboard controls. - Improves hold-drag Backspace selection and release-to-delete behavior. - Adds foreground recording controls and opt-in, content-free diagnostic logging.
  • Sep 12, 2026 0.1.20
    - Keeps the APK compact with clear, verified on-demand Parakeet model setup - Improves the true-black voice-input interface and model guidance - Replaces the launcher and store artwork with a circular black icon and white microphone
  • Sep 11, 2026 0.1.19
    - Introduces the Offline Speech to Text identity and logo - Enables private, caller-supplied audio transcription for app integrations