Android v0.15.0: a new local voice architecture
Loquela v0.15.0 introduces Batch, Streaming, and Two-pass recognition modes while improving privacy guidance, diagnostics, vocabulary import, and Enter behavior.
- Version
- 0.15.0 (27)
- Channel
- Google Play open testing
- Status
- Available on Google Play open testing
Loquela v0.15.0, version code 27, is now available through Google Play open testing in Australia, the United Kingdom, and the United States.
Loquela v0.15.0 contains the largest change to its local voice system since downloadable speech models were introduced. The app can now describe and run different recognition plans instead of treating every local model as the same kind of batch recognizer.
Existing models and ordinary offline dictation remain supported. Streaming and Two-pass modes appear only when the app has a compatible, validated model, so this architecture can grow without silently changing a user's selected model or forcing two large downloads onto every phone.
Added
- Setup now explains Android's standard third-party keyboard warning and links to Loquela's privacy details.
- Advanced diagnostics can retain the last 5, 10, 25, or 50 transcription failure reasons without storing dictated text or audio.
- Personal Dictionary can import words from a text or CSV file or from Android's user dictionary after a private on-device preview.
- Advanced voice diagnostics include an optional Parakeet tail-padding experiment for investigating clipped final words.
- Local recognition now supports three explicit session plans: Batch, Streaming, and Two-pass.
- Batch keeps Loquela's existing behavior. The selected offline model receives the complete utterance after Stop and returns one final transcript.
- Streaming feeds a native online model while the user speaks, allowing replaceable partial text to appear in the active editor before the utterance ends.
- Two-pass pairs one streaming model with one batch model. The first model supplies immediate partial text, and the second performs the authoritative final decode after Stop.
Changed
- Enter inserts a newline in multiline editors and honors Search, Go, Send, Next, Done, and Previous actions in single-line action fields.
- Paid settings no longer carry repeated Free or Full labels. Locked controls explain that Full is required and link to the upgrade screen.
- Local model packages receive stronger validation and safer repair behavior when an install is missing or damaged.
- The local speech runtime has been upgraded without replacing existing batch models.
- Recognition modes appear only when a compatible installed or accessible model can provide them.
Fixed
- Local recognition now preserves the final partial audio buffer instead of dropping it during handoff.
- A missing or incompatible selected model is shown as needing repair instead of silently switching recognition modes.
- Local model recovery now checks that the complete expected package survived an interrupted update or cancellation.
Known issues
- New Moonshine v2 and Nemotron streaming packages remain hidden until they pass physical-device accuracy, speed, memory, battery, thermal, and editor-safety checks.
- Streaming must stay ahead of incoming audio on a supported phone, and Two-pass must provide a meaningful final correction before either mode is promoted.
- History can record a send action performed by Loquela, but Android does not let the keyboard observe an app's own Send button.
- The Parakeet tail-padding control is diagnostic and remains off by default while clipped-ending reports are measured.
For the deeper engineering story, read How Loquela is building live and two-pass local voice typing.
Get Loquela or visit Support.