Skip to content

FUTO Voice Input Moonshine v1.4.2-beta.7

Pre-release
Pre-release

Choose a tag to compare

View all tags
@github-actions github-actions released this 28 Aug 11:45
· 11 commits to master since this release
v1.4.2-beta.7
7796555

Beta 7 highlights

  • Adds optional on-device final-transcript cleanup with S1-mini by Superwhisper.
  • Clearly labels the feature English-only and bypasses cleanup unless English is established.
  • Adds styling, structure, context, model warm-time, and runtime controls.
  • Downloads and verifies the pinned Q4_K_M model at runtime; model weights are not bundled in the APK.
  • Benchmarks CPU thread counts and experimental OpenCL on the phone, selecting OpenCL only when its validated result is at least 15% faster.
  • Runs cleanup in an isolated app process with a 15-second raw-transcript fallback.
  • Adds transcript-free performance diagnostics with copy and ZIP export.

What to test

  • Download S1-mini from Model Options, enable it, and dictate English on a Galaxy S24+.
  • Try every styling, structure, and context option and confirm cleanup runs only after Stop.
  • Try non-English and mixed-language configurations; confirm raw text is retained when English is not established.
  • Compare Auto, CPU, and OpenCL and rerun the optimization test while watching latency and thermals.
  • Test immediate unload and each keep-warm duration, cancellation, long dictation, and low-memory recovery.
  • Export the diagnostics ZIP after any slow or incorrect run and attach it with feedback.

Known beta gaps

  • This beta has not yet been performance-validated on the target Galaxy S24+.
  • OpenCL support depends on the phone vendor's driver and automatically falls back to validated CPU inference.
  • Cleanup may return the raw transcript after a timeout, model/process failure, or memory pressure.
  • Release-level thermal, memory, long-transcript, and full backend regression testing remains in progress.

Installation and models

This is a signed standalone APK for the FUTO Voice Input fork. Moonshine remains the default offline recognizer; Parakeet, Nemotron, and legacy Whisper/GGML remain selectable in Model Options.

Speech models and S1-mini weights are not bundled in the APK. Download the desired model on first use or from Model Options; transcription and cleanup remain offline after installation.

The Parakeet integration and repository changes in this fork were built with AI assistance using Codex (GPT-5).

Full Changelog: v1.4.2-beta.6...v1.4.2-beta.7

Assets 3
Loading