ScribeKey 0.9.0 for Android
What changed in ScribeKey 0.9.0 for Android, released 2 August 2026.
2 August 2026. All releases.
First build handed to beta testers, through Firebase App Distribution.
Added
- Transcribe an audio file you already have: pick a file from Home, and the recording is kept, badged in history and retryable like any other
- The Dictionary now shows the corrections ScribeKey taught itself, and lets you delete one or clear them all
- Prompts can be pointed at the apps that use them, so email gets the email prompt
- Interrupted model downloads resume where they stopped, and free space is checked before one starts
- Failed and cancelled dictations stay in history with their audio, so you can retry them
- Spoken punctuation can be switched on, with a haptic to confirm it landed
Changed
- First run rebuilt around the deck pattern: one decision a screen, stated as its consequence
- The cloud screen now owns the connection, not the capability; each pipeline stage chooses its own placement
- Rewriting settled on one name, one writer and an amount that reads as a scale
- Every surface got an edge, and cards and buttons share one language
- Home leads with time transcribed, measured from real recordings
- The Settings hub lost its longer sections and now counts setup honestly
- Rewriting stages are skipped below a word threshold and abandoned on a timeout
- The app id moved from the legacy package ID to
app.scribekey - Added an About ScribeKey seven-tap unlock for Advanced Labs; experimental switches are managed in that single destination
- Gated HF Cloud, Audio Scribe, the Audio Scribe model-library route, its import control, benchmark tools, and stale onboarding selections behind the Labs feature gate
- Simplified normal Voice Engine and Enhance menus by removing one-option controls and redundant explanatory copy
- Updated Settings, About/Labs, Voice Engine, and Enhance wording for faster scanning
- Made Labs unlock-progress snackbars replace the current message instead of queuing older banners
- Updated M3 Expressive screenshot baselines for the reduced menu surfaces
Fixed
- Importing long audio no longer runs out of memory
- Transcripts no longer arrive with a leading space
- A custom prompt is no longer overridden by the system template, and its safety rules survive
- The dictation rate is weighted by how long each session ran, so one short fast burst no longer skews it
- Failed dictations no longer count towards time transcribed
- Cleanup strips only hesitation sounds, and honours the custom filler list
- Retry on Home actually retries, and a failure keeps you in the app
- Switching the cloud service away from OpenAI kept the OpenAI endpoint
- The overlay no longer blames the microphone for every failed dictation
- Stored HF Cloud or Audio Scribe selections now resolve to the safe local speech path while their Labs feature is unavailable
- Settings search and diagnostics no longer expose benchmark tools before the benchmark Lab is enabled
Removed
- The engine redesign prototype and its Labs entry