v0.3.0-alpha — onboarding, consent-first capture, low-end support

Muesli for Android — third alpha

Tested end-to-end on a Samsung Galaxy S20 FE and a low-end POCO C71 (Android 15). Sideload the APK (arm64, Android 7.0+), then follow the new onboarding — it walks you through model download and a test dictation.

New since v0.2.0-alpha

• iOS-style onboarding — Profile (name + use case) → Microphone → Keyboard → Speech Model (with download + %/MB progress) → Try It (real first dictation) → AI Summaries (optional). New installs can no longer land in the app unable to transcribe.
• Consent-first capture — the keyboard never records on its own: keep-mic-ready now prewarms instead of auto-recording, and the old auto-stop-on-pause system recognizer (which could also send audio to Google) has been removed entirely, along with the mock sandbox. On-device Parakeet is the only engine.
• Parakeet 110M recommended for fresh installs — validated on low-end hardware (POCO C71): fast and accurate; the 600M v3 remains for faster phones.
• Meeting audio playback — play/pause + seek for retained meeting WAVs in the detail view.
• Recording microphone picker — Automatic / Phone / Bluetooth / USB, with live route label on the hero pill and a Settings card (iOS AudioInputRouteManager parity).
• Low-end resilience — fixed OOM on ~256MB-heap devices (append-only audio buffering), partial-decode backlog guard for slow chips, and HuggingFace DNS blocks now fall back to a mirror with faster failure.
• Polish — vector Muesli mark everywhere, faithful monochrome notification glyph, expandable voice-note previews (Show more/less), slimmer quick-capture capsule, natural notification copy.

Known gaps

• No crash recovery or sync yet; Bluetooth mic routing varies by OEM
• Release APK signed with the debug key; arm64 only

Full changelog: v0.2.0-alpha...v0.3.0-alpha

v0.2.0-alpha — quick capture, model catalog, diarization

Muesli for Android — second alpha

Big feature drop on top of the first alpha. Same install: sideload the APK (arm64, Android 7.0+), transcription model downloads in-app.

New since v0.1.0-alpha

• Quick capture (Android-exclusive) — a draggable floating bubble over other apps: tap to dictate into a slim floating capsule (stop · waveform · elapsed · cancel); the note is saved and copied to your clipboard. Also a Quick Settings tile and home-screen widget. Enable in Settings → Voice Notes → Quick Capture.
• Model catalog — pick between Parakeet v3 600M (multilingual, default) and the new Parakeet 110M English model (126MB, CTC) in Settings → Models; per-model download/delete/switch.
• Silero VAD meetings — meetings chunk on real speech boundaries instead of fixed 30s windows.
• Speaker diarization — finished meetings are diarized fully on-device (bundled pyannote + TitaNet models) and transcripts render grouped by speaker (Speaker 1 · 0:04). Toggle in Settings → Meetings.
• Design polish pass — deliberate letter-spacing across the type scale, tabular stats, 12dp buttons.
• Expandable voice notes — Show more/Show less on long dictations.
• Vector Muesli mark — crisp traced logo everywhere; proper monochrome notification glyph; natural notification copy.
• About screen now shows the real version; UI copy reflects the selected model.

Known gaps (unchanged)

• No meeting playback, crash recovery, or sync yet
• ChatGPT sign-in flow still needs a real-account test
• Release APK signed with the debug key; arm64 only

Full changelog: v0.1.0-alpha...v0.2.0-alpha

v0.1.0-alpha — first alpha

Muesli for Android — first alpha

First installable alpha of the Muesli Android port (parity-in-progress with muesli-ios).

What's in

• Voice Notes home: stats (streak / words / WPM / meetings), hero record card with live waveform, recents list
• On-device transcription: sherpa-onnx + Parakeet TDT 0.6B v3 (int8), resumable ~638MB model download
• Muesli keyboard (IME): iOS-style layout, waveform strip, punctuation row, press-hold repeat
• Meetings end-to-end: foreground recorder service, 30s chunked decoding, WAV retention, 7 summary templates, transcript/notes tabs, search
• AI summaries: ChatGPT sign-in (OAuth PKCE) primary + OpenRouter BYOK fallback, auto-titles
• Settings hub: Voice Notes / Meetings / Dictionary / Models / AI Summaries / Appearance (dark/light + blue/green/slate accents) / About
• Official Muesli icon + iOS design tokens, launch warmup screen

Known gaps

• No live transcript preview in meetings (deliberate)
• ChatGPT sign-in flow untested with a real account yet
• No diarization/speakers, VAD chunking, WAV playback, or model catalog yet
• Signed with the debug key; arm64 devices only

Install

Download the APK below and sideload (arm64 Android 7.0+). Transcription model downloads on first use over Wi‑Fi.