Parlando 0.5.46 (build 225), dictation for Mac by Epiphani Studios. Last updated Saturday, November 14, 2026. Every row names the mechanism and a way to check it. Requirements: macOS 15 or later on Apple silicon (check the downloaded app with vtool -show-build for minos 15.0 and lipo -archs for arm64).
| Capability | How it works | Default | Verify |
|---|---|---|---|
| Hold-to-talk dictation into any app | A listen-only key monitor watches one modifier key; on release, the finished text is inserted at the focused text field through the macOS Accessibility API. | Right Option. Right Command or Globe selectable. | Hold the key in any text field, speak, release. |
| Hands-free latch | Tap the key (under 0.30 s), then press again within 0.45 s; dictation continues until the next press. | On | The panel shows "Latched". |
| On-device speech recognition | Apple SpeechAnalyzer on macOS 26 and later with two modules: DictationTranscriber biased by your vocabulary ("Personal") or SpeechTranscriber ("General"). On macOS 15 through 25 it is Apple's on-device speech recognition. | Personal | Menu bar > Speech model. |
| Accuracy pass | When you let go of the key, a second model re-reads the same audio and replaces the live text when it passes a check against it. Engines: On this Mac (Whisper large-v3), Cloud (OpenAI's newest transcription model with your own key, $10 default monthly limit), or Automatic (the cloud when the Mac is busy or warm, this Mac otherwise). The pass and the cleanup steps share one finishing deadline, 1.9 seconds on the Mac, and the best checked text at the deadline is typed. An optional setting types dictations under 12 seconds the moment you let go. | On this Mac | Settings > Accuracy > Engine. |
| Streaming panel | A non-activating floating panel shows partial words and a level meter; it never takes keyboard focus. | On, Medium, bottom of screen | Settings > While you are speaking. |
| Text insertion | Automatic: up to 200 characters typed as keystrokes; longer text pasted, with the clipboard restored about 0.35 s later. Always-type and always-paste selectable. | Automatic | Menu bar > Insert text by. |
| Smart spacing and capitalization | Reads at most 12 characters before the cursor at insertion time to choose leading space and capitalization; keeps "I", acronyms, camelCase, and dictionary words. | On | Dictate mid-sentence and at the start of a line. |
| Cleanup | Removes whole-word fillers (um, uh, erm, uhm, hmm, mm, mhm, er, ah), capitalizes the first word, closes sentences at the transcriber's pauses with a period or a question mark, and applies "new line" and "new paragraph". | Light | History shows what you said beside what was typed. |
| Recovery when nothing is focused | If no text field has focus, nothing is typed; a card shows the text with a Copy button, and the dictation is saved to history. | On | Click the desktop, then dictate. |
| Cancel and Undo | Escape cancels a dictation while you are speaking. An Undo button brings it back for six seconds, and an X puts the notice away. | On | Press Escape while dictating. |
| Language | US English by default; macOS installs a missing speech model on first use through its AssetInventory service. | en-US | Run Parlando --locales to list installed models. |
| Capability | How it works | Default | Verify |
|---|---|---|---|
| Recognition vocabulary | Plain words (names, brands) are passed to the Personal module as contextual strings, biasing recognition before it decides. | Empty until added | Dictionary pane, Words filter. |
| Corrections | "Heard, written" pairs applied on whole words, longest phrase first, case-insensitive unless flagged; "at Name" becomes "@Name". | Empty until added | Dictionary pane, Corrections filter. |
| Snippets | Say a short trigger phrase and the stored long text is typed. | None until added | Snippets pane. |
| Import from a file | One entry per line: heard -> written, comma or tab separated, or a bare word; lines starting with # or // skipped. | Manual | Dictionary > Import. |
| Import from Wispr Flow | Reads ~/Library/Application Support/Wispr Flow/flow.sqlite read-only and merges corrections, vocabulary, snippets, and mentions, keeping existing entries and listing every skipped row with a reason. | Manual | Menu bar > Import from Wispr Flow. |
| Style by kind of writing | Four surfaces (personal messages, work messages, email, other) each take a tone: Formal, Casual, very casual (personal only), or Excited!. Style sets punctuation and capitalization, keeps your words as you said them, and runs on the Mac. | Formal everywhere | Style pane, one tab per surface. |
| Per-app rules | Per application bundle: fillers, spoken commands, capitalize-first, leading space, and smart spacing, each inheriting the global setting when unset. Suggested rules offered for Claude, ChatGPT, Slack, Messages, and Mail. | No rules | Settings > Per-app rules. |
| Capability | How it works | Default | Verify |
|---|---|---|---|
| Transforms on a key | Select text in any app and press Option plus 1 to 8; the selection is replaced by the model's result, or left untouched if anything fails. Built in: Polish, Prompt Engineer, Bullets, Slack, Email, Professional, Strip the AI, My voice; create your own instruction and key. | Keys bound; needs a key to run | Transforms pane. |
| Providers | Anthropic (api.anthropic.com/v1/messages) or OpenAI (api.openai.com/v1/chat/completions) with the user's own API key and chosen model; the key is stored in ~/.parlando/cloud.json with file mode 0600. | No provider set | Settings > Cloud, Test it. |
| Cloud repunctuation | Sends finished transcript text for punctuation. It uses the result when every word remains as you said it, or types the on-device text instead. | Off | Settings > Cleanup engine. |
| Medium cleanup and re-process | Rewrites for clarity and concision through the chosen provider; the original is kept and can be restored. | Off (Light is default) | Settings > Auto cleanup; Dictations row menu. |
| Polished cleanup | The newest Claude Haiku turns run-on dictation into clean sentences and keeps every number, negation, dictionary name, and capitalized name, or types the original text instead. Monthly limit: $3. | Off (Light is the default level) | Settings > Auto cleanup. |
| Capability | How it works | Default | Verify |
|---|---|---|---|
| History | Each dictation (text, destination app bundle id, duration, word count, delivered or not, and raw text when different) appended to ~/.parlando/history.jsonl. | Kept 24 hours; three days, seven days, or forever selectable | Dictations pane; Settings > History. |
| Insights | Dictations, words, minutes, words per minute, day streak, a per-day chart, and a per-app breakdown, computed from local history. | Available | Insights pane. |
| Menu bar recents | The 12 newest dictations, click to copy; undelivered ones marked. | Available | Menu bar > Recent dictations. |
| Model Context Protocol (MCP) server | JSON-RPC 2.0 over stdio, started by an assistant with Parlando --mcp. Nine tools: get_preferences, set_preference, add_substitutions, add_vocabulary, remove_entries, set_app_rule, list_app_rules, remove_app_rule, search_history. | Runs only when a client starts it | Add the binary path with --mcp to an MCP client. |
| Usage counters | Local counters over a closed list of event ids plus total seconds and characters, in ~/.parlando/stats.json on the Mac. | On | Parlando --stats. |
| Capability | How it works | Verify | |
|---|---|---|---|
| Signed, notarized download | Developer ID Application: JASON PAEZ (34MSZG2KRW), hardened runtime, notarized, and stapled; the disk image is signed and notarized too. | spctl -a -vv /Applications/Parlando.app reports "Notarized Developer ID". | |
| Automatic updates | Reads https://updates.belstone.com/parlando/appcast.json about 20 s after launch and every 6 hours; verifies SHA-256, a strict deep code signature, and Team ID before installing; waits for any dictation in progress. | Menu bar > Check for Updates now. | |
| Permissions | Microphone and Accessibility to dictate, plus Screen Recording when meeting notes capture the other side of a call and its shared slides; entitlement com.apple.security.device.audio-input. | System Settings > Privacy & Security. | |
| Audio path build check | The build refuses to compile if any networking symbol appears in the six source files that capture, transcribe, format, or insert dictation. | Source build script, network-symbol gate. | |
| Latency build check | Every build runs the shipped binary on a 2.79-second recording and refuses to finish if the tail from end of speech to final text exceeds 350 milliseconds (a regression tripwire, far above the measured 0.06 to 0.09 seconds). | On every build | Source build.sh, step "latency regression". |
| Measurement | Result | Conditions |
|---|---|---|
| Speech engine final text after a 2.79 s recording ends | 0.061, 0.067, 0.072, 0.092 s | Parlando 0.2.0 (111), --bench at real-time pace, MacBook Pro M1 Max, macOS 26.6.2, load average 85 on 10 cores, October 4, 2026 |
| First partial words, same runs | 0.535 to 0.551 s | Same |
| Speech engine final text after a 14.96 s recording | 0.167, 0.172, 0.175, 0.198 s | Same |
| Proper nouns, one six-name sentence | Personal 5 of 6; General 2 of 6 | 452-word personal vocabulary, M5 MacBook Air, August 3, 2026, as Detto |
| Proper nouns, 40 real recordings (37 names) | Personal 30 (81%); General 25 (68%) | 456-word personal vocabulary, one speaker, August 4, 2026, as Detto |
Download: getparlando.com/download (the current build, read from the same feed the app updates from). Machine-readable: /download/latest.json. Previously named Detto; dettodictation.com redirects here.