Velora 0.8.0

Velora 0.8.0

Local-first, on-device dictation for macOS. Your voice never leaves your Mac.

New in this release

• Speaker diarization for meetings. Multi-person calls are split into
  Speaker 1 / Speaker 2 / … in the transcript, using on-device models
  (sherpa-onnx, ~46 MB fetched on the first meeting). One-on-one calls stay
  plain "Them"; toggle in Settings → Meetings.
• Safe Voice Edit. Select text anywhere, press ⌥⇧E, and speak an edit —
  "make this more formal", "fix the grammar", "turn this into bullet points".
  Only the selection is touched, the change is undoable (⌘Z), and the edited
  text is always on the clipboard as a fallback.
• "As Heard" escape hatch. Paste the untouched raw transcript from the
  menubar (Reformat Last as → As Heard) or view it in History — for when
  cleanup got something wrong and you just want the words.
• Automatic update checks. A once-a-day anonymous check against GitHub
  releases (off in Settings → General if you'd rather Velora never touch the
  network after model download).
• Menubar icons for History, Meetings, and Setup Assistant.

Performance

Cleanup model is auto-selected by RAM (lighter 4-bit on ≤16 GB Macs).
Measured on Apple Silicon: dictation stop→final ~1.1 s, voice edit ~0.5 s per
sentence, diarization ~2 s per audio-minute. Peak engine memory a few GB.

Install

Apple Silicon + macOS 14+. Grab the DMG below (Developer ID-signed and
notarized), drag Velora to Applications, and open it. First launch downloads
the models (~6 GB) with visible progress; after that everything is offline.

Prefer to build? git clone, then make app — see the README.

Velora 0.6.0 — Voice Intelligence, Meeting Memory, and Local Agents

Velora 0.6.0 — voice intelligence, meeting memory, and local agents

This release makes Velora useful beyond the moment text is inserted, while keeping the same local-first boundary.

What’s new

• Voice Intelligence: see words and dictations over time, estimated time saved, streaks, app/mode usage, latency, cleanup rate, and an honest zero-edit rate with observation coverage. Share cards contain aggregate numbers only.
• Private Meeting Memory: get optional suggestions for Zoom, Google Meet, Teams, Slack Huddles, and nearby Calendar events. Recording always requires explicit confirmation. Velora keeps microphone and computer audio as separate local tracks, then creates a searchable transcript, summary, decisions, and action items.
• Local CLI and MCP: opt in to an owner-only local control socket for history, search, stats, file transcription, and one visibly approved live dictation. There is no network listener, and access is off by default.

Under the hood

• Meeting post-processing is chunked and resumable; live dictation takes priority.
• Active meeting capture is finalized on graceful quit. Microphone audio uses a crash-resilient CAF container; after a crash or forced termination, frames that reached disk are preserved for explicit retry instead of being orphaned.
• Interrupted processing resumes, while permanently failed meetings wait for a manual retry and repeated engine restarts are capped to prevent poison-file loops.
• Meeting audio has configurable local retention and can be played, exported, retried, or deleted from Settings.
• Intelligence was tested against a 100,000-row history, including migration from older databases.
• History/search responses return only the user-visible final text and bounded display metadata; they omit the raw-vs-cleaned transcript pair, bundle identifiers, audio paths, screen context, contacts, learning data, and internal row/session identifiers. File transcription returns only its caller-supplied path, local timing metadata, and cleaned text.
• Long-running CLI/MCP requests are bounded and cancelled when their requesting client disconnects; that cancellation is scoped to the exact request. Quitting also revokes pending consent, cancels active agent work, and closes accepted control clients before engine teardown.

Upgrade

Replace the existing app in Applications with the new DMG. Existing preferences, history, modes, dictionary entries, and archived dictation audio are preserved. Local CLI/agent access and Calendar matching both remain off until you enable them.

Velora 0.4.5 — faster dictation and better punctuation

Velora 0.4.5 makes dictation faster without swapping or shrinking either configured model.

Performance

• Reuses the exact Qwen prompt prefix prepared while you speak instead of paying that cost after release.
• Cancels stale preview and cleanup work, and restarts the sidecar if MLX inference hard-wedges.
• Stops repeated full-recording audio copies during Whisper previews.
• Keeps long Terminal dictations streaming even when the first segment is short.
• Pauses hidden HUD animations; the installed app and loaded sidecar measured 0.0% idle CPU across 12 samples.

Writing quality

• Complete thoughts now receive sentence-ending punctuation.
• Longer Terminal prose keeps its final full stop while short commands remain model-free and command-safe.
• Qwen fixes clear grammar and agreement issues conservatively without paraphrasing or changing meaning.
• Names, numbers, technical terms, and dictated details stay protected by the divergence checks.

Live transcript

• The HUD shows up to two readable lines selected on word and sentence boundaries.
• Whisper previews use the same warmed model that produces the final transcript and cannot corrupt final state.

The production models remain mlx-community/whisper-large-v3-turbo and mlx-community/Qwen3.5-4B-MLX-8bit.

Verified with 233 Python tests, 69 Swift self-checks, six exact-model quality fixtures, an installed end-to-end audio run (1.298s stop-to-final), Developer ID signing, Apple notarization, stapling, and Gatekeeper assessment.

SHA-256: a5541af5be8570573fc9f416ec167a4a7737da650f578ec25780d78e627c1671

Velora 0.4.4 — reliable model setup and notarized install

Velora 0.4.4 gives first-run model setup an honest, reliable onboarding flow and installs normally through macOS Gatekeeper.

First-run setup

• The Try It field stays locked while the speech and writing models complete their one-time setup.
• Onboarding shows the current setup phase and download percentage.
• Setup progress remains correct across engine reconnects.
• Superseded pre-ready connections can no longer send duplicate ready frames or prevent setup completion.
• If the optional writing model cannot load, Velora falls back to raw dictation instead of trapping onboarding.

Distribution

• The app and DMG use Developer ID signing and hardened runtime.
• Apple notarization is stapled and verified before publication.
• Gatekeeper accepts both the disk image and bundled app as Notarized Developer ID.

Verified with 55 Swift self-checks and 199 Python tests.

SHA-256: ef5641094743bcdb1c10ed6e9af01f4ba73f56a273a3681a50d599dfb3632374

Velora 0.4.3 — notarized install and honest first-run setup

Velora 0.4.3 is the first public Developer ID-signed and Apple-notarized release. Drag it to Applications and open it normally; no Gatekeeper bypass is needed.

First-run setup

• The Try It field stays locked while the speech and writing models complete their one-time setup.
• Onboarding shows the current setup phase and download percentage instead of letting a first dictation fail mysteriously.
• Progress and completion remain correct across engine reconnects.
• If the optional writing model cannot load, Velora falls back to raw dictation instead of trapping onboarding.

Distribution

• The app and DMG use Developer ID signing and hardened runtime.
• Apple notarization is stapled to the DMG and verified before publication.
• Gatekeeper accepts both the disk image and bundled app as Notarized Developer ID.

Verified with 55 Swift self-checks and 198 Python tests.

SHA-256: 2cf998df0d728cfe4ceae9b6f88196ed90794b0a451b14683ac3087350fe0de8

Velora 0.4.1 — visible first-run setup (and we're open source now!)

The repository is now public — MIT-licensed, secret-scanned across the full history.

First-run experience

A fresh install now shows exactly what's happening while Velora sets itself up (previously a silent multi-minute wait):

• Onboarding — the try-it step shows a live progress card: "Downloading the speech model (1.6 GB) — 42%", then unlocks dictation the moment the model is ready.
• Menubar — a live-ticking setup line + tooltip while models download.
• Hotkey — dictating too early tells you the phase and percentage instead of a vague "starting…".
• Speech recognition unlocks first (~1.6 GB); AI cleanup follows in the background (~4.3 GB). Everything downloads once from Hugging Face, then Velora is fully offline.

Verified end-to-end on a simulated fresh machine (empty caches, real downloads).

Installing: drag Velora to Applications. The app is self-signed (not yet notarized), so the first open needs right-click → Open. Models download automatically on first launch — watch the menubar or the onboarding window.

Velora 0.4.0 — file transcription, voice commands, big audit round

New

• Transcribe Audio File… (menubar): drop in a voice memo or meeting recording — transcribed locally in chunks with live progress, result copied to the clipboard and saved as a <name> transcript.txt next to the audio. A live dictation always takes priority; the job pauses and resumes around it.
• Voice commands v1: say just "scratch that" / "undo that" to undo the last dictation, "new line" / "press enter" for Return, "new paragraph" for a blank line. Whole-utterance only — command words inside normal dictation are never intercepted. Toggle in Settings → Dictation.
• Usage stats in History: words today / all-time, estimated time saved vs typing, and your daily streak.
• Personal dictionary import/export (Settings → Dictation): move learned corrections + vocabulary between Macs as JSON.

Faster

• final no longer waits for the audio archive write — FLAC encoding moved off the stop→final path.
• Fixed a potential multi-second stall at dictation start when the target app is busy (AX timeout).

More reliable (cross-vendor audit round — all P1 findings fixed)

• Switching audio devices mid-recording (AirPods connecting, mic unplugged) no longer silently kills capture — the recording just continues on the new device.
• Model switches refused by a busy engine no longer leave the Settings picker lying; engine state is authoritative.
• Fixed a config write race that could silently revert settings changed during a model download.
• History rows can no longer be silently dropped when background vocab mining reads the DB (sqlite busy timeout).
• Learned corrections: a one-off content edit (e.g. vercel→Netlify) can no longer become an instant global rewrite; genuine mishearings still learn on first sight.
• Undo command is keyboard-layout-safe (no ⌘W surprises on AZERTY), pastes are protected on slow Electron apps, plus a dozen smaller fixes.

196 engine tests + 49 app self-checks green. All local, zero network, as always.

Velora 0.3.1 — Smartness v2

Smartness v2 — the self-learning release

Spoken self-corrections, done right. "Let's meet at 3pm — no no, 6pm" → Let's meet at 6 p.m. Works for any cue — "oh sorry, Shubham", "I mean Rahul", "scratch that", "delete this line" — across sentence boundaries, driven by the LLM's understanding of speech repair (a replacement swaps a parallel item: a name for a name, a time for a time). Emphasis ("no no no!"), real apologies, and content "no"s are preserved.

Real-time learning loop (all local). Edit the text Velora just pasted and it learns your fix within seconds — the HUD pill returns with the mishearing struck out → your correction → Learned. Misheard names commit on the first fix; everything is manageable (and deletable) in Settings.

It mines your vocabulary while idle. When nothing is happening, the local LLM extracts names and jargon from your dictation history into an auto-vocabulary that biases recognition itself (whisper initial_prompt) — "Velora" stops coming out as "valora" at the STT level.

Flat latency at any length. Long dictations are decoded and cleaned in pause-aligned segments while you speak: a 75-second dictation finishes ~1.5s after you stop (previously fell back to raw, unpunctuated text) — with live partial transcript in the HUD, now also for whisper.

Terminals got smart. Long prose dictated into a terminal (Claude Code, codex chats) is cleaned and punctuated; short shell commands stay verbatim.

Everything runs on-device. Zero network. Zero telemetry.

180 engine tests; adversarially reviewed (2 independent models, 8 findings fixed); real-model eval 15/16 on the self-correction suite.

Velora 0.2.0

On-device dictation with smarter cleanup and controls.

Fixed

• Spoken punctuation — "full stop", "question mark", "comma", etc. now become symbols instead of leaking as literal text (with guards so real prose like "came to a full stop" is preserved).
• Cleanup timeout now scales with length — long paragraphs no longer silently skip cleanup.

New

• Code mode has a real AI instruction; terminals get their own verbatim mode (commands stay untouched).
• Smart model selection by your Mac's RAM (compact / balanced / quality tiers) + a cleanup-model picker.
• Model storage manager — see each cached model's size and reclaim space from ones you don't use.
• Draggable HUD — drag the pill anywhere; position is remembered.
• Manageable learned corrections — view and delete what Velora learned from your edits.
• Auto-versioning on every build.

Both parakeet speech models (higher-quality streaming + cheaper English-only) remain available alongside the whisper default.

---

Local-first: your voice never leaves your Mac. Self-signed (not notarized) — right-click → Open on first launch. Models download from Hugging Face on first use.

Velora 0.1.0

First public build of Velora — local-first, on-device dictation for macOS (Apple Silicon).

Install

1. Download Velora-0.1.0.dmg below, open it, drag Velora to Applications.
2. This build is self-signed, not yet notarized, so the first open needs right-click → Open (or System Settings → Privacy & Security → Open Anyway).
3. On first launch, grant Microphone, Input Monitoring, and Accessibility when prompted — all three are required for the hotkey and text insertion to work.
4. The engine downloads its models (~4 GB) from Hugging Face on first run; after that everything runs offline.

Prefer to build from source? See the README — make app.

Highlights

• Multilingual dictation (English, Indian English, Hindi + more) via on-device whisper-large-v3-turbo
• App-aware LLM cleanup with custom modes/prompts
• History browser, 6-month audio archive + reprocess, optional romanized (Hinglish) output
• Zero network calls at dictation time

⚠️ Requires Apple Silicon (M1+) and macOS 14+.