What happened
The Phone Whisper repository describes Audio To Text as an Android utility for transcribing audio shared from other apps, using local models through sherpa-onnx. Its changelog says version 1.2.0, dated October 17, added a Live tab between Catalog and Info. The tab transcribes the device microphone in real time using a streaming model. Pressing Start opens the microphone and displays recognized text in a scrollable area; the control changes to Stop. The tab is enabled only when the selected model supports streaming, with Kroko given as an example.
The release also added a local archive for live sessions that produce text. Entries are saved automatically to a local SQLite database, with periodic autosave approximately every 10 seconds and a final save on stop. The repository says this means an abrupt process kill loses at most the last autosave interval. The archive lists each session’s date, time, and duration in minutes and seconds. Users can review the full transcript, share it as plain text, or delete it with confirmation. A live session and selected tab survive configuration changes such as rotation.
Version 1.2.1, dated October 24, changed session behavior and added interface cues. During a live session, the screen stays awake until the user presses Stop, without an additional permission; normal screen timeout resumes afterward. The microphone is used only while the Live screen is visible. Locking the device, pressing Home, or switching apps ends the session and saves text produced up to that point. The foreground microphone service and persistent notification with a Stop action were removed. When the user returns after an interruption, a dismissible notice appears; it does not cover the transcript or reappear after the app is closed. Navigation to Home, Catalog, or More info is blocked during recording, with Stop as the way out. The release also says model lists now mark whether each model supports live transcription.
What the archive adds
New Runtime covered Phone-whisper as an Android app for local voice transcription on October 31, 2026. This is a new fact on a known story: the project has since added streaming microphone transcription and a local session archive, then clarified foreground-only recording behavior.
Trend signal
This update supports generative media becoming application infrastructure: transcription is integrated into a mobile workflow with model selection, live capture, persistence, and sharing, rather than remaining a one-off model interaction. The evidence is a small project’s implementation, not proof of broad adoption or a general market shift.
Why now
The repository dates the live-transcription release to October 17 and the recording-behavior update to October 24, 2026. The changelog does not explain why the changes were timed this way, so there is no supported broader timing claim.
Why it matters
Mobile transcription has two distinct operational needs: keeping capture local and making a session’s state legible when the user leaves the app. This implementation pairs on-device inference with a visible recording screen, a deliberate Stop action, and periodic persistence. That combination offers a useful design reference for teams building voice features that should not silently continue recording in the background.
The tradeoff is equally concrete. Leaving the app ends capture, so the design favors a clear foreground boundary over uninterrupted recording. A local archive also shifts responsibility to the device: users can inspect, share, or delete transcripts there, while teams still need to decide how long such records should remain and how deletion behaves across backups or exports. The project changelog does not answer those broader product-policy questions.
Limits and uncertainty
The repository describes capabilities and implementation behavior, but does not report transcription accuracy, latency, battery use, supported device performance, or comparative results across models. It names Kroko as an example of a streaming model, but does not quantify which models work well for which languages or environments. The stated autosave interval bounds expected loss from process termination, but does not establish durability under every storage failure. These claims are not independently tested here, and the repository’s small footprint means the release should be read as a project update rather than evidence of a widely adopted approach.
Skill to build now
For any foreground voice feature, write down the session contract before implementation: what starts capture, what stops it, what happens on app backgrounding or screen lock, and how much work an interruption can lose. Then test those transitions explicitly on-device. A periodic save interval and a final save on Stop are concrete mechanisms to consider when transcripts must survive process interruption.
Reader takeaway
Phone Whisper’s update is less about a new transcription model than about the surrounding product contract: supported streaming models enable live capture, transcripts are saved locally, and leaving the visible screen ends the session. Builders can borrow that explicit boundary and autosave pattern, while treating quality, latency, and device costs as unanswered until measured.