Show HN: Velora – On-device macOS dictation (Whisper and a local LLM, no cloud)

Developer Anurag Roy released Velora, an open-source, MIT-licensed macOS dictation tool that runs entirely on-device using Apple Silicon's MLX framework, with no cloud dependencies. The app uses Whisper for speech-to-text and a local LLM for cleanup, ensuring audio and transcripts never leave the user's Mac, and offers app-aware formatting and a private meeting memory feature.

Local-first dictation for macOS. Hold a key, speak, release — polished text appears in whatever app you're using. Every dictation step, from speech-to-text to AI cleanup, runs on-device via MLX https://github.com/ml-explore/mlx on Apple Silicon. Audio, transcripts, and history never leave your Mac; confirmed Personal Dictionary terms can sync privately through your iCloud Drive. Dictation tools like Superwhisper and Wispr Flow proved the product: invisible, fast, smart voice input everywhere. But Wispr Flow transcribes in the cloud, and both are closed-source subscriptions. Velora is the open-source answer: Private by architecture, not by toggle. Zero network calls at dictation time. Models are downloaded once from Hugging Face; after that the engine never touches the network. No accounts, no telemetry, no analytics. Free and open. MIT-licensed. Modes, prompts, and vocabulary are plain JSON files you own. Fast enough to trust. Speech-to-text streams while you talk , so text lands moments after you release the key — AI cleanup included. Doesn't rewrite you. Cleanup is deliberately conservative: transcribe-don't-answer, no added content, and a divergence guard that falls back to your raw words if the LLM drifts. The raw transcript is always kept in history. Hold-to-talk and toggle dictation — hold Right-Option default to record, release to insert; or double-tap / click the menubar icon to toggle. Esc always cancels cleanly. Hotkey behavior is configurable. Capsule HUD — a small floating capsule with a live 24-bar waveform from your mic. It never steals focus, and morphs through listening → transcribing → inserted/error states. Start/stop sounds included toggleable . Multilingual — the default whisper-large-v3-turbo model handles English, Indian English, Hindi, and dozens more languages with automatic language detection. Non-Latin transcripts Devanagari, CJK, Arabic skip the English-tuned cleanup so they stay faithful. Optional romanized output — a Dictation-settings toggle transliterates non-Latin speech into Latin letters Hindi → natural Hinglish, e.g. "नमस्ते आज मौसम" → "Namaste aaj mausam" , keeping English words English. Off by default. App-aware smart formatting — Velora detects the frontmost app and picks a mode automatically: Slack / Messages / Discord / Telegram / WhatsApp → terse, casual, no trailing period on a single short sentence Mail → professional structure and tone Notes / Obsidian / Notion / Bear → markdown allowed, lists when you enumerate VS Code / Cursor / Terminal / iTerm / Ghostty / Warp / Zed → raw mode, no AI rewriting - everything else → clean, well-punctuated default On-device STT + LLM — transcription with whisper-large-v3-turbo multilingual; parakeet-tdt-0.6b-v2 available for streaming English, plus a Hindi/Hinglish specialist and cleanup/formatting with Qwen3-4B-Instruct-2507-4bit . Cleanup removes fillers, applies self-corrections "no wait, I meant Tuesday" , punctuates, and honors spoken "new line" / "new paragraph". History browser — every dictation raw + final text, app, mode, duration is stored in a local SQLite database. Browse, search, copy, or paste-again from the History tab; the menubar menu shows your last three click to copy . Voice Intelligence — an on-device dashboard turns local history into useful trends: words and dictations over time, estimated time saved, streaks, app and mode breakdowns, STT/cleanup latency, cleanup rate, and an honest zero-edit rate with observation coverage. Share cards contain aggregate numbers only — never transcript, app, or contact text. Private Meeting Memory — Velora can suggest recording when Zoom, Google Meet, Teams, or a Slack Huddle is active, with optional Calendar matching. Capture starts only after an explicit confirmation, keeps microphone "Me" and computer audio "Them" as separate local tracks, and produces a searchable transcript, summary, decisions, and action items in the background. Processing is resumable and live dictation takes priority. Speaker diarization — when more than one person is on the other side of a call, the transcript splits them into Speaker 1 / Speaker 2 / … using on-device diarization sherpa-onnx: pyannote segmentation + titanet embeddings, ~46 MB downloaded on the first meeting, sha256-pinned . One-on-one calls stay plain "Them"; any diarization failure falls back cleanly. Toggle in Settings → Meetings. Safe Voice Edit — select text in any app, press ⌥⇧E configurable , and speak an edit: "make this more formal", "fix the grammar", "turn this into bullet points". Only the selection is touched, the result replaces it in place, and ⌘Z undoes it. The edit prompt is benchmarked 94% on a 50-command suite, spikes/engine/bench voice edit.py and guarded — an unusable result keeps your text unchanged, and the edited text is always on the clipboard as backup. "As Heard" escape hatch — when cleanup gets something wrong, paste the untouched raw transcript from the menubar Reformat Last as → As Heard or view it in History. No re-processing, works even after the audio clip has been pruned. Local CLI + MCP — scripts and local agents can inspect allow-listed history/stats, transcribe an audio file, or request one visibly approved live dictation. Access is off by default, runs through an owner-only Unix socket instead of a network server, and never exposes raw audio, screen context, contacts, or learning data. Personal Dictionary — teach Velora exact names, product terms, and optional “heard as → write as” corrections. Edit-learned and auto-learned words stay visible and reversible, and only confirmed dictionary entries sync through your app-specific iCloud Drive folder when iCloud is available. Audio archive + reprocess — clips are saved as compact FLAC under ~/.velora/audio configurable retention, default 6 months / 4 GB cap so you can re-transcribe any past dictation with a better model straight from the History tab. Custom modes editor — every mode is a JSON file in ~/.velora/modes/ , editable from the Modes tab: per-mode LLM prompt the Superwhisper-style feature , formatting level, app bindings, vocabulary, and replacements. Drop in a file to create your own see Customization customization . Model picker — choose your STT model in Settings from the engine's registry, with managed downloads. Live spectrum waveform — the HUD's 24-bar waveform is driven by a real FFT of your mic, so bars react to both loudness and pitch. Safe insertion — clipboard is snapshotted and restored around the synthesized ⌘V, with a keystroke-typing fallback for apps that block paste. Secure input fields passwords are detected and insertion is suppressed. - Apple Silicon Mac M1 or later - macOS 14+ uv https://docs.astral.sh/uv/ manages the Python engine - Swift toolchain — Command Line Tools are enough xcode-select --install ; no Xcode needed - ~4.4 GB of disk for the default models downloaded on first run 1. Prerequisites one-time xcode-select --install Swift toolchain Command Line Tools curl -LsSf https://astral.sh/uv/install.sh | sh uv, for the Python engine 2. Build & run git clone https://github.com/sushilk1991/velora cd velora make app builds build/Velora.app SwiftPM release + hand-rolled dev bundle open build/Velora.app make app compiles the Swift app and bundles the Python engine; the engine's dependencies are fetched by uv on first launch. First launch then walks you through onboarding: microphone permission, accessibility permission live-detected as you grant it , hotkey choice, and a try-it playground — you finish with a real dictation. The engine downloads the default models from Hugging Face on first run ~6 GB; live progress shows in the onboarding window, the menubar menu, and the HUD if you try dictating early — speech recognition unlocks first, AI cleanup a few minutes later . After that, everything is offline. Prefer not to build? Install with Homebrew: brew install --cask sushilk1991/tap/velora Or grab the latest release from Releases https://github.com/sushilk1991/velora/releases , drag Velora to Applications, and open it. Run the .app bundle, not the bare binary — macOS permission grants mic, accessibility attach to the signed bundle identity. Tests: make test runs the Python engine suite. The app also has an embedded Swift self-test swift run Velora --selftest , and make perf-test checks Intelligence against a 100,000-row history. The installed app includes a CLI at: /Applications/Velora.app/Contents/Resources/bin/velora --help Enable Allow local CLI and agents in Settings → General, then use status , recent , search , stats , transcribe , or listen . Add --json for machine-readable output. listen always displays an approval prompt before the microphone starts. Velora also exposes the same narrow surface as a local MCP stdio server: /Applications/Velora.app/Contents/Resources/bin/velora mcp The app must be running. Nothing listens on the network, and disabling the setting immediately removes history, stats, and action access; status remains available so tools can explain what is missing. The cleanup model is picked by RAM tier at first launch Qwen3-1.7B-8bit on ≤14 GB Macs, Qwen3-4B-4bit up to 24 GB, Qwen3.5-4B-8bit above and can be changed in Settings → Models. Measured on Apple Silicon M-series: | Metric | Measured | |---|---| Streaming STT throughput parakeet-tdt-0.6b-v2 | ~184× realtime | Multilingual transcription whisper-large-v3-turbo , default | ~0.3 s per clip | | LLM cleanup, warm | ~0.5–0.8 s per paragraph | | Voice edit selection + spoken instruction | ~0.4 s per sentence, ~1 s per paragraph | | Meeting diarization | ~2 s per audio-minute, ~0.5 GB peak | The default multilingual model is batch transcribes on release rather than streaming , a deliberate quality tradeoff for accurate Hindi/Indian-English. For streaming English with live HUD partials, pick parakeet-tdt-0.6b-v2 in Settings. If cleanup would blow its budget, Velora inserts the raw transcript immediately instead of making you wait — the cleaned version is never the bottleneck. Two processes, one product: ┌────────────── Velora.app Swift, SwiftPM ──────────────┐ │ menubar · hotkeys · mic/HUD · insertion · settings │ │ history/intelligence · meeting capture/store · consent │ └────────────┬───────────────────────────────┬──────────────┘ │ engine.sock │ control.sock │ framed JSON + PCM │ owner-only JSON ┌────────────┴────────────────────┐ ┌─────┴─────────────┐ │ velora-engine Python + MLX │ │ local CLI / MCP │ │ STT · cleanup · meeting notes │ │ default-off │ └─────────────────────────────────┘ └───────────────────┘ The Swift app owns everything user-facing; the Python engine is an invisible inference server, supervised spawned, health-checked, restarted by the app. Full details, wire protocol, and module map: docs/ARCHITECTURE.md /sushilk1991/velora/blob/main/docs/ARCHITECTURE.md . Modes live in ~/.velora/modes/ .json built-ins are copied there on first run — edit away . The file format is the API: { "name": "Standup", "prompt": "The user is dictating a daily standup update. Keep it to short bullet points grouped under 'Yesterday', 'Today', and 'Blockers' when the speech covers them.", "formatting": "full", "apps": "com.tinyspeck.slackmacgap" , "vocabulary": "Velora", "MLX", "parakeet", "Kubernetes" , "replacements": { "vs code": "VS Code", "k eight s": "k8s" } } prompt — mode-specific instructions merged into the cleanup system prompt. formatting — "off" regex-level tidy only, no LLM , "light" , or "full" . apps — bundle ids that auto-activate this mode when frontmost. An explicit mode selection always wins over app matching. vocabulary — proper nouns and jargon hinted to transcription and cleanup. replacements — literal text substitutions applied after cleanup. Edits are picked up via the engine's config reload — no restart dance required. - Models are downloaded once from Hugging Face. Model downloads, Personal Dictionary iCloud Drive sync, and an optional update check are the only network-backed features; Velora has no backend service. - The update check is one anonymous HTTPS GET to the public GitHub releases feed, at most once a day, carrying nothing about you or your dictations. Turn it off in Settings → General and Velora never touches the network at all after model download. - At dictation time there are zero network calls — audio, transcripts, and cleaned text never leave the machine. - History is a local SQLite file under ~/.velora/ . Delete it whenever you like. - Meeting audio, transcripts, and notes live separately under ~/.velora/meetings/ . Meeting detection uses local app/window metadata and, only if enabled, nearby Calendar events. Detection can suggest a recording but can never start one; every recording requires an explicit confirmation and shows a persistent menu-bar indicator. - Intelligence share cards are rendered locally from fixed labels and numeric aggregates. They cannot include transcript, app, contact, or calendar text. - Local agent access is off by default. When enabled, only processes running as your macOS user can reach the owner-only ~/.velora/control.sock ; there is no TCP listener. Live microphone use still requires approval for each request. - Personal Dictionary sync uses your standard iCloud Drive protection and contains only confirmed terms and corrections — never audio, transcripts, history, screen context, or model data. It remains fully usable offline and does not use a Velora server. - No accounts, no telemetry, no analytics. This is enforced by architecture, not by a settings checkbox. Velora is actively developed and pre-1.0. The complete local dictation loop now sits alongside Voice Intelligence, consent-first Meeting Memory, and the default-off CLI/MCP surface. The release gate combines the Python engine suite, hundreds of embedded Swift checks, a 100,000-row Intelligence benchmark, packaged-app CLI/MCP smoke tests, and Apple signing/notarization verification. Known limitations: The — microphone and accessibility TCC grants attach to the signed bundle, so hotkeys and insertion won't work from a bare .app bundle is required for real use swift run binary. Batch default — the multilingual default model transcribes on release, not live; switch to parakeet-tdt-0.6b-v2 for streaming HUD partials English only . Speaker labels are acoustic, not identities — "Me" is the microphone track; remote voices are clustered into "Speaker 1/2/…" by how they sound. Velora never claims to know who a speaker is. Voice-edit casing quirks — capitalization-specific instructions "fix the weird capitalization" are the 4B model's weakest edit category; grammar, tone, shortening, and list edits are reliable. Contributions welcome — see CONTRIBUTING.md /sushilk1991/velora/blob/main/CONTRIBUTING.md for dev setup, repo layout, and PR guidelines. Building the app locally is just make app ; no signing setup or Apple credentials are needed. Licensed under the MIT License /sushilk1991/velora/blob/main/LICENSE .