Loquela
Patch Notes

Android v0.15.1: local live voice typing enters beta

Loquela v0.15.1 opens Nemotron-powered Sequential and Both voice modes to beta testers, adds smaller Moonshine v2 models, and fixes local audio transfer failures.

2 min read
Version
0.15.1 (28)
Channel
Google Play open testing
Status
Available on Google Play open testing

Loquela v0.15.1, version code 28, is now available through Google Play open testing in Australia, the United Kingdom, and the United States.

This focused beta follow-up puts the new local voice architecture into testers' hands. It makes the first genuine on-device live model available, exposes all three recognition choices, and fixes the audio-transfer failure found on the Play-installed v0.15.0 build.

Added

  • Nemotron 3.5 ASR Streaming 0.6B is available as Loquela's first genuine live on-device voice model.
  • Voice settings now always show Batch, Sequential, and Both.
  • Sequential displays replaceable text while you speak using Nemotron.
  • Both uses Nemotron for live text and a compatible batch model for the authoritative final transcript after Stop.
  • Selecting Sequential or Both without Nemotron explains what is missing and links directly to Voice Models.
  • Moonshine v2 Tiny and Base add substantially smaller English batch-model choices.
  • Moonshine v2 Tiny is a free low-end model choice while its beta performance is evaluated.
  • Voice Models exposes a Parakeet ending test for investigating clipped final words.
  • General Settings now links directly to Loquela's feature-request, bug-report, and feedback page.

Changed

  • The older, larger Moonshine v1 packages are no longer offered for new installs; existing installs remain usable and receive upgrade guidance.
  • A missing or damaged model keeps the selected recognition mode visible and explains how to repair it instead of silently switching modes.
  • Nemotron and Moonshine v2 are clearly marked as beta validation work rather than being hidden from the people testing the beta.

Fixed

  • Local dictation no longer fails because the isolated recognizer receives an incomplete audio stream.
  • Automatic Both-mode pairing can choose a compatible installed streaming model and batch finalizer when explicit slots have not been assigned.

Known issues

  • Nemotron's speed, memory, battery, temperature, and editor behavior still need broader physical-device testing.
  • Both requires Nemotron plus a compatible installed batch model and costs more storage, memory, and processing than either pass alone.
  • The Parakeet ending test is off by default because its best padding value has not yet been established across devices and utterances.

For the design and tradeoffs behind these modes, read How Loquela is building live and two-pass local voice typing.

Get Loquela, browse Voice Models, or send feedback.