Dictation and voice

Hold ⌘⇧D (Ctrl+Shift+D on Windows and Linux), speak, and release it. Wrengle cleans up what you said and drops it into the note, assistant draft, Plugin Builder prompt, or chart text field you had open. Holding the shortcut always dictates in Quick Dictation, the plain-speech mode, no matter which mode you have saved.

For a longer session, press ⌘⇧V / Ctrl+Shift+V to start listening, and press it again — or click Dictate / Voice in the status bar — when you're done. A toggled session keeps listening until you stop it, rather than only while you hold the key.

Wrengle has two voice modes, chosen under Settings → Dictation:

• Quick Dictation (Recommended) turns everything you say into text for the open note, assistant draft, Plugin Builder prompt, or chart text field. Trigger words such as "editor" or "terminal" are just ordinary words in this mode — only a short list of exact session and correction phrases act as controls.
• Voice Control also dictates plain speech, but adds trigger words — editor, workspace, agent, terminal — that run commands instead of typing them out.

Voice Control is the default mode, and Wrengle remembers your last choice between sessions. Dictation cleanup starts on Light edit, and response speed starts on Fast; switch to Quick Dictation under Settings whenever you want a dictation-only path.

Where your voice goes

Speech recognition and voice analysis are two separate settings under Settings → Dictation, and each one resolves to a local or a cloud route on its own:

• Speech recognition turns your voice into text. Local speech recognition (Whisper) keeps your microphone audio on your device. Cloud speech recognition — OpenAI, Deepgram, or ElevenLabs — streams your microphone audio to whichever provider you chose.
• Voice analysis cleans up that text and, in Voice Control, turns it into commands through your selected direct OpenAI/Anthropic model or Wrengle AI tier. On Wrengle AI, analysis text and replies transit Wrengle and OpenAI. For Voice Control commands, analysis can send the current transcribed utterance, the open note's title, up to 40 folder paths, and up to 15 recent note paths. For Light edit, it sends the current utterance and up to four recent dictated sentences from the same voice session, all captured within one exact target: the note, assistant draft, Plugin Builder prompt, or chart text field. While one in-progress phrase and its exact target remain active, Light edit may analyze provisional transcript snapshots that can later be superseded or cancelled, but no more often than once every 1.5 seconds. Each request can use provider quota or billing, or Wrengle AI credits. Turning live preview off suppresses those during-speech calls; final cleanup still runs. It does not send that target's identity, your microphone audio, or separately read or attach arbitrary typed text, the note body, or surrounding chart content; its payload is bounded to those dictated slots.

Cloud analysis can reach the network even when speech recognition stays local. The two settings resolve independently, so picking local speech alone does not keep a session entirely on your device — check the analysis setting too if that matters to you.

Both settings default to Auto, but they use separate credentials and fallback rules. Speech Auto can fall back to Local Whisper. Analysis Auto is cloud-only: it prefers the most recently connected configured OpenAI or Anthropic AI provider, then checks OpenAI and Anthropic, and uses that provider's reviewed default model. Without a usable AI key, analysis pauses for setup instead of falling back to a local model.

A confirmed key is not the same as a working key. Auto only checks that Wrengle can read a key from your keychain — it does not contact the provider to confirm the key still works. An invalid, revoked, or over-quota key can pass that check and still fail once you actually start speaking.

Choosing Local, a specific provider, or turning Cloud voice features off overrides Auto, and that choice sticks even as you add, replace, or remove keys. Resetting a speech or analysis preference sends it back to Auto.

Starting a voice session

The Dictate or Voice control in the status bar starts and stops the selected mode the same way the toggle shortcut does, if you'd rather click than remember a key combination.

While a session is active, the heads-up panel shows what's happening: session state, your microphone level, any device errors, the live transcript, recent results, and where the words are headed — a note name, Assistant draft, Plugin Builder prompt, or a chart field such as Node label or Page name. If no destination is shown, click into an editable note, the assistant composer, a Plugin Builder prompt, or a chart text field before you keep talking.

Stopping — with the button, the toggle shortcut, or by releasing Hold to Dictate — commits whatever is pending, even if live preview is off. Wrengle stays briefly in a stopping state while it finishes cleanup and inserts the words.

A spoken stop can drop your last few words; a button or shortcut stop cannot. Saying "stop listening," "stop dictating," or "stop voice" ends the session without committing an unfinalized live partial. Use the button or the keyboard shortcut instead when you need to keep the very last thing you said.

Sometimes Wrengle can't insert right away — a confirmation is open, an agent preview is showing, focus is missing, the assistant composer, Plugin Builder prompt, or chart field changed, or you switched dictation targets. When that happens, it holds your pending text in an expandable recovery card:

• Insert here puts it into whatever is focused now.
• Return to original reopens the note it was meant for, if that was an editor — focus that editor, then choose Insert here.
• Copy puts it on your clipboard.
• Discard clears it without inserting anything.

That recovery text stays scoped to the vault it came from. Switch vaults and Wrengle discards it, so it can never land in a same-named note somewhere else.

After dictation lands in a note, assistant draft, Plugin Builder prompt, or chart text field, Undo last phrase reverts the latest finalized or checkpointed dictation change — as long as that exact target is still valid for undo. Accepted corrections to up to four recent dictated sentences are applied atomically. The current unfinalized phrase remains grey ghost text and is not part of that undo step yet.

Chart dictation works in node labels, connector labels, swimlane lane headers, and page names. Focus the field before starting with the same status-bar control or shortcut used for notes. Node labels accept paragraph breaks; connector labels, lane headers, and page names stay single-line and reject them.

Voice and meeting recording share one capture path: you can't start a voice session while a meeting is recording or finishing up, and you can't start a meeting recording mid voice-session either.

Speech recognition

In-app dictation has its own speech setting, separate from meeting transcription: Settings → Dictation → Dictation speech engine.

Auto works through that list in order: the provider whose key you connected or verified most recently, then — if that one isn't available, or you're on an older install with no history yet — OpenAI, Deepgram, and ElevenLabs, before it falls back to Local Whisper. A key Wrengle can't read doesn't count.

A cloud choice needs two things: the matching key under Settings → Dictation → Speech keys, and Cloud voice features turned on in Settings → Privacy & Data. That access defaults on when you've never set it, so saving a connected key is all Auto needs. Turn it off once, explicitly, and it stays off — Wrengle then keeps automatic speech on Local Whisper and blocks generative analysis. Pick a specific cloud provider and later lose access, and that choice stays selected but unusable until you restore access.

Speech keys are separate from your AI-chat keys. Wrengle asks Deepgram to opt your account out of model-improvement use, which Deepgram says can affect pricing. ElevenLabs realtime dictation follows whatever logging and retention your ElevenLabs account already has — an ordinary BYOK account doesn't get enterprise Zero Retention Mode from Wrengle. Provider billing, logging, and retention terms apply regardless.

Turning cloud access off stops active capture right away; any audio or analysis already in flight gets a short window to finish.

Deepgram and ElevenLabs receive your microphone audio for as long as their cloud stream is active. OpenAI is different: dictation stays gated on your device. Wrengle trims leading silence, keeps up to 240 ms of audio before each detected speech window so the first sound is never clipped, sends that short lead-in with the window, and commits each window by hand. Turning up response speed does not turn this into one continuous stream to OpenAI.

If a cloud stream disconnects, Wrengle keeps the session open, leaves your last ghost text visible, shows a reconnecting state, and keeps buffering your microphone audio while it retries. If that buffer fills up before reconnecting, Wrengle tells you some speech may be missing.

Dictation language defaults to English. Choose Auto detection, English, English (US), English (UK), Spanish, French, German, Italian, Portuguese, Portuguese (Brazil), Dutch, Polish, Japanese, Korean, Chinese, or Hindi. It only changes what Wrengle listens for — the interface, the built-in control phrases, and the command examples all stay in English this release. An English-only Whisper rung can't serve a non-English or Auto selection, so install a multilingual one for those.

Personal vocabulary feeds names, acronyms, and specialist terms to the speech engine as hints, not guarantees — it improves recognition but doesn't promise a particular word gets used. Local Whisper uses them on your device. Deepgram and ElevenLabs receive them in each provider's own key-term format when you pick that provider. ElevenLabs' current keyterm guide caps realtime keyterms at 50 entries of 20 characters each and lists an extra charge for them; Wrengle only sends entries that fit. OpenAI Realtime Whisper doesn't use personal vocabulary this release.

Voice analysis

Voice analysis & polish model defaults to Automatic. Automatic uses the most recently connected configured OpenAI or Anthropic provider, then OpenAI and Anthropic in that order, with the provider's reviewed default model; it does not select Wrengle AI. There is no local generative fallback. Direct and enabled managed models remain available as sticky explicit choices. If an explicitly selected provider's key later goes missing, Wrengle keeps that route selected and reports the missing credential instead of switching destinations. Reset returns to Automatic.

Quick Dictation with Verbatim cleanup skips this setting entirely. Voice Control needs it to run commands, and Light edit needs it to clean up your wording and, when useful, revise up to four recent dictated sentences from the same session and target. Continuous Light edit checkpoints can make repeated requests and use provider billing or Wrengle AI credits, subject to the 1.5-second minimum interval described below.

Microphone and voice setup

Settings → Dictation → Microphone and voice setup puts your speech engine's readiness next to your microphone controls:

• Dictation microphone picks a listed input, or System default microphone. Use Refresh microphones after you plug in or remove hardware.
• Test microphone runs a five-second level check on the selected input — nothing more. It does not start speech recognition, analysis, or a cloud request. Watch the meter and device label to confirm Wrengle hears you; choose Stop test to end early.
• Speech engine ready means your selected or Auto-resolved path — Local Whisper or cloud speech — is set up. Speech engine needs setup points at what's missing: a speech model, a provider key, or disabled cloud voice access.

A test and a dictation or meeting recording can't both hold the microphone at once. If a microphone you specifically selected disappears, Wrengle reports the error instead of silently switching to another device — reconnect it, pick a different input, or fall back to System default microphone. Because the current audio layer doesn't expose stable hardware IDs, Wrengle also leaves identically named microphones out of persistent selection rather than risk a silent wrong choice — use the system default, or keep only one of those devices connected at a time.

Quick Dictation

Quick Dictation sends ordinary recognized speech straight to wherever you're working. Words like editor, workspace, agent, and terminal are just words here — nothing classifies your prose into a command. This is the simplest path for taking notes or drafting an assistant prompt.

Saying "stop listening," "stop dictating," or "stop voice" as a whole utterance ends the session immediately, without running it through trigger or command analysis. Say those words in the middle of a longer sentence and they stay literal text. "Scratch that" and "undo that" remove your latest eligible dictation the same way; if there's nothing eligible to undo, Wrengle tells you instead of deleting something else.

Spoken punctuation is off by default. Turn it on and the exact standalone words comma, period, full stop, question mark, colon, semicolon (or semi colon), open quote, close quote, and new paragraph insert that punctuation or break instead of being typed out. Wrengle only catches these as whole utterances, so saying "a question mark" inside a longer sentence just types those words.

Voice Control

Every command needs a trigger word — the word that tells Wrengle which part of the app you're talking to. Speech without one is plain prose, dictated into whatever you're focused on.

While a confirmation card is open, say yes, yeah, yep, confirm, do it, ok, or okay to accept it, and no, nope, cancel, or never mind to reject it. A choice card also takes first through fourth, or the name of a visible option. These only count as whole replies — Wrengle does not pull them out of ordinary dictation.

Spoken approval never runs a shell command or sends a request to a hosted cloud or external agent. Those cards need a click, or keyboard activation, instead — so your microphone can't approve its own request to cross that boundary. Spoken cancel still works on them.

When the assistant composer or Plugin Builder prompt is focused, dictation drafts a prompt there and never sends it — review or edit it, then send with Enter or the Send button. Trigger words become part of that draft instead of commands: saying "agent summarize this" drafts "summarize this," and "terminal run tests" drafts "run tests" rather than running anything. Say "literally" before a trigger word to dictate the word itself.

Voice commands

Say a trigger word, then describe what you want: editor, workspace, agent, or terminal. Trigger words are yours to change under Settings → Dictation → Voice trigger words — set a global word for each area, or a different one for the editor than for the terminal, so each area answers to whatever feels natural to you.

editor

Structure the open note without touching the keyboard.

• Headings: "editor heading two Project Goals" makes a level-2 heading reading "Project Goals."
• Lists: "editor bullet list milk eggs bread" makes a three-item bullet list — items split on natural pauses, "and," or "next." "editor numbered list" and "editor checklist" work the same way.
• Blocks: "editor quote," "editor divider," "editor code block," "editor callout," "editor task," and more — the same blocks in the slash menu.
• Breaks: "editor new line" for a line break; "editor new paragraph" for a new block.
• Shape: "editor indent" or "editor outdent" to nest a block; "editor make this a heading two" to convert one; "editor delete block"; "editor undo."

Format and refine text

Apply formatting and fix small mistakes without leaving voice.

• Format what you just said: "editor bold that" bolds the text you just dictated. "editor italic that," "editor strikethrough that," and "editor code that" work the same way.
• Format named words: "editor bold the word budget" emphasizes one word; "editor italic from quarterly to review" formats everything between two words you name.
• Clear formatting: "editor clear formatting" strips bold, italic, strikethrough, and inline code from your selection or the words you name.
• Fix the last few words: "editor delete last word" removes the word you just said; "editor delete last sentence" removes the sentence.

Every one of these is reversible — say "editor undo," or use the editor's own undo, to step back. The same formatting is also available as buttons in the editor's toolbar, for when you'd rather use your hands.

Say "literally" before a trigger word to dictate it as text — "literally editor," for example.

workspace

Controls notes, folders, navigation, and settings.

• Notes: "workspace new note called Budget" · "workspace open my budget note" · "workspace rename this note to Q3 Budget" · "workspace delete the meeting note" (asks you to confirm).
• Folders: "workspace new folder called Projects" · "workspace delete the archive folder" (asks you to confirm).
• Search: "workspace search quarterly review" · "workspace search tax documents."
• App actions: "workspace toggle terminal" · "workspace toggle assistant" · "workspace focus mode" · "workspace open settings" · "workspace theme dark" (also light, system).

agent

Hands your request to whichever assistant is currently selected or resolved by Auto — a built-in OpenAI or Anthropic assistant, or an external agent you've configured. That assistant's own context, permission, provider, authentication, and billing boundaries still apply.

Outside the focused composer, every built-in cloud or external-agent request shows you the full request in a confirmation card first and waits for a click or keyboard approval. When the composer itself is focused, the same trigger word just drafts your prompt instead, so you can speak naturally into it before sending.

• "agent summarize this note"
• "agent write a packing list for the trip"
• "agent rewrite this section more concisely"

terminal

Controls the built-in terminal. A run command needs a terminal already open, rejects hidden or control characters, and shows you the exact command in a confirmation card first — only a click or keyboard approval sends it to the shell.

• Shell writes (need a click or keyboard approval): "terminal run npm test" · "terminal run git status" · "terminal clear."
• Session control: "terminal open" · "terminal close" · "terminal toggle."
• Tab navigation: "terminal next tab" · "terminal previous tab."

Safety tiers

Voice actions sit at different safety levels:

• Read-only and view actions — open, search, toggle a panel, change the theme — run right away.
• Creating or renaming a note or folder also runs right away, with an Undo action right there so a mistake is one step from fixed. Say "undo" at any point to reverse your last voice action.
• Deleting a note or folder always asks first. Wrengle shows a confirmation card naming the target; say "yes" or "cancel," or click. One pending confirmation has to resolve before another can replace it.
• Agent requests always need your confirmation first. A built-in cloud or external agent may send request and note context under that backend's own privacy boundary.
• Terminal run commands need a running terminal and your confirmation, then execute in the embedded shell with normal OS permissions. They can change files, reach other services, or have other side effects — session controls like open, close, and toggle are not confirmation-gated.

Voice Undo cannot pull back a sent agent request or an executed terminal command. Once one of those leaves Wrengle, whatever recovery the destination itself offers is the only way to walk it back.

Cleanup and response speed

Configure this under Settings → Dictation:

• Verbatim cleanup keeps your exact wording, removing only speech and pause artifacts.
• Light edit removes fillers and false starts, fixes casing and punctuation, and keeps your intended wording. It is contextual editing, not summarisation: it works on the current utterance and can revise up to four recent dictated sentences from the same voice session, all captured within one exact target: the note, assistant draft, Plugin Builder prompt, or chart text field. It never rewrites arbitrary note text or surrounding chart content.
• Response speed — Fast, Balanced, or Deliberate — sets how long Wrengle waits before committing a phrase, for every speech engine: Local Whisper, OpenAI, Deepgram, and ElevenLabs. Fast is the default and commits after a short pause, so it can split your speech more often; Balanced gives a little more thinking time; and Deliberate waits longest before committing. During one in-progress phrase on an unchanged target, Light edit may analyze provisional transcript snapshots repeatedly, but never more often than once every 1.5 seconds. A snapshot can later be superseded or cancelled, and each analysis can use provider quota or billing, or Wrengle AI credits.
• Dictation language and Personal vocabulary guide recognition; Spoken punctuation turns on the exact Quick Dictation punctuation phrases.
• Show live dictation preview controls whether your words show up as grey ghost text at the note's caret. Turn it off to keep raw words in the voice panel and pause during-speech Light edit analysis; final cleanup still runs.

Quick Dictation with Verbatim skips analysis completely — it never loads, warms, or calls a generative model. Voice Control still needs analysis for commands, and Light edit still needs it for cleanup.

Grey preview text isn't saved until it settles. A complete thought settles sooner than a fragment trailing off on "and" or "because," so you can keep going mid-thought without Wrengle jumping ahead.

The current unfinalized phrase stays in the grey ghost preview while Light edit runs. Accepted corrections to eligible earlier sentences land atomically as one undoable change; the current phrase is inserted only when it finalizes. Moving the caret, switching targets or sessions, crossing a paragraph boundary, or editing eligible text stops Light edit from revising those earlier sentences.

A long pause can settle the current phrase, but it does not by itself clear eligible prior dictation. Light edit can still revise those sentences when you continue in the same active voice session and exact target without moving the caret, crossing a paragraph boundary, or editing the eligible text.

Stop with the button or shortcut, and Wrengle freezes the pending words, runs your chosen cleanup, and inserts the result before the session goes idle. If cleanup takes too long or fails, it inserts a cleaned-up raw version instead of dropping your words and leaves earlier sentences untouched.

Switch notes, change editor focus, or move to a different assistant composer before the preview settles, and Wrengle will not insert it into the new surface. That pending text stays with the note or composer it was meant for.

Cleanup never turns your words into structure — markdown-looking dictation stays literal text. Saying - item, # heading, bold, italic,

code, or label without a trigger inserts those exact characters as prose, not a list, heading, bold or italic text, inline code, or a link. Use voice commands with a trigger word to actually structure or format: with Spoken punctuation off (the default), saying "new paragraph" just types those two words; turn Spoken punctuation on and that exact phrase makes a break instead. In Voice Control, "editor new paragraph" always makes one. Likewise, "bold budget" as plain prose types those words, while "editor bold budget" applies bold formatting.

Privacy and requirements

Your speech and analysis choices resolve independently — here's what each combination needs and sends:

For cloud Voice Control command analysis, bounded context can include the active note's title, up to 40 folder paths, and up to 15 recent note paths. Cloud Light edit instead receives the current utterance plus, when they are still eligible, up to four recent dictated sentences from the same voice session, all captured within one exact target: the note, assistant draft, Plugin Builder prompt, or chart text field. While one in-progress phrase and its exact target remain active, it may analyze provisional transcript snapshots that can later be superseded or cancelled, but no more often than once every 1.5 seconds. Each request can use provider quota or billing, or Wrengle AI credits. Turning live preview off suppresses those during-speech calls; final cleanup still runs. It may return revisions to those sentences alongside the cleaned current utterance. Wrengle applies accepted corrections to earlier dictated sentences as one undoable atomic change, while the current unfinalized phrase remains grey ghost text until finalization. Light edit does not send that target's identity or separately read or attach arbitrary typed text, the note body, or surrounding chart content; its payload is bounded to those dictated slots.

Quick Dictation with Verbatim cleanup is the one exception to that analysis column: it only needs your chosen speech path, no analysis model at all. Voice Control needs analysis even with Verbatim cleanup on.

The speech and analysis columns can use different providers. Choose Auto speech with an ElevenLabs transcription key while Automatic analysis resolves OpenAI, for example, and your microphone audio goes to ElevenLabs while the resulting text goes to OpenAI. Turning Cloud voice features off forces automatic speech back to Local Whisper and blocks generative voice analysis; Verbatim Quick Dictation still works without analysis. It doesn't touch your separately resolved assistant. Meeting live captions and meeting transcription keep their own local-first defaults; cloud meeting final-pass transcription remains a separate opt-in of its own.

Wrengle keeps a small set of performance marks for the current session only — capture-ready, first transcript, speech-end-to-final, insertion, and stopping timing. Each mark holds just a stage name and an elapsed time: no audio, no transcript or utterance text, no note paths or titles, no provider hosts, no keys. None of it goes to remote telemetry, and all of it disappears when you quit Wrengle.

Voice sessions don't use meeting recovery storage. The temporary .app/live/ records the meeting docs describe exist only for meeting recording, finalizing, saving, report generation, and recovery — they hold finished transcript records and exact save/report bindings, never raw audio or the in-progress prepared report.

A request you hand off to a hosted assistant follows that assistant's own privacy rules — the same as if you had typed it into the assistant panel yourself.

Setup and troubleshooting

• Open Settings → Dictation and choose your mode, speech engine, language, cleanup level, response speed, personal vocabulary, spoken punctuation, and trigger words. Speech and analysis start on Auto; choose an explicit direct OpenAI/Anthropic model or enabled Wrengle AI tier only when you want to override Automatic.
• For local speech, install the Whisper model Wrengle offers when you first start a session. For cloud speech, save and verify the matching OpenAI, Deepgram, or ElevenLabs key — Auto uses it right away unless you've explicitly turned Cloud voice features off.
• A specific local Whisper choice always uses that exact downloaded, language-compatible rung, never a silent substitute. Auto picks the fastest installed rung that fits. After a local dictation, Wrengle keeps that model warm for about 60 seconds so a quick restart is faster.
• Voice Control and Light edit need a usable direct OpenAI/Anthropic key or a selected, funded Wrengle AI route. Automatic picks only a configured provider key and its reviewed model; without a usable route, you can still use Quick Dictation with Verbatim cleanup, and Wrengle preserves the raw transcript when analysis is unavailable.
• On macOS, Wrengle asks for microphone access during startup. If you denied it, or revoked it later, use Settings → Privacy & Data → Desktop access to open the operating-system settings, then retry or restart as directed there.

On the current unpackaged Windows build, capture permission always shows as Not required. That build cannot yet reliably request or report per-app capture consent, so it does not probe the device or the global desktop-app switch at startup. If Windows or your selected device is actually blocking capture, the later device-open attempt fails prompt-free instead of showing a permission dialog — the Privacy page always offers a Windows microphone Settings link so you can fix it there.

On any platform, Test microphone and the dictation shortcuts never open a permission prompt of their own. Use Dictation microphone to choose a device, then Test microphone to confirm its level. If a saved device goes missing, choose Refresh microphones and pick a connected input, or fall back to System default microphone.

• Focus an editable note, the assistant composer, a Plugin Builder prompt, or a supported chart text field, and check its name in the destination indicator. Wrengle blocks a target change while text is pending, rather than inserting it into the wrong surface.
• Stop or finish any meeting recording or finalization before you start dictating — meeting and voice capture can't share the microphone at the same time.
• If cloud dictation shows Reconnecting, leave the session open while it retries. A long outage can fill the audio buffer Wrengle keeps for this; any speech that gets lost is reported rather than silently treated as transcribed.

Limitations

Recognition quality depends on your microphone, your environment, the language you picked, any vocabulary hints, and the transcription model itself. Picking a recognition language does not translate Wrengle's interface or its built-in English control phrases.

In Voice Control, only an exact trigger match runs a command — an unrecognized one surfaces in the panel instead of quietly succeeding. Quick Dictation never treats a trigger word as a command, no matter what you say.

Confirmation gates a delete, a shell command, or any agent request before it runs — it does not make the result reversible afterward. Once one of those actions changes something outside Wrengle, voice Undo generally cannot undo it.

Voice capture is macOS-first, following the same platform support as meeting capture.