LLM polish
Polish organizes ordinary voice typing after recognition and before output. It is off by default. Enable it under Settings → Text processing → LLM polish, then configure a provider and model. It is separate from the voice-command intent model and can share the same provider’s Keychain credentials.
Styles & customization
Section titled “Styles & customization”| Style | Behavior |
|---|---|
| Verbatim | Preserve wording and fix clear recognition errors |
| Light polish | Remove fillers and self-corrections with light formatting |
| Deep polish | Produce clearer written language while retaining meaning |
With polish enabled, add glossary terms or custom instructions and use the page’s test area to check the output. Trial / Pro offers app-aware smart polish: light for chat, deep for writing/mail, and verbatim for coding/terminals. Unmatched apps use the baseline style.
Trial / Pro also offers Rewrite when text is selected, which turns ordinary voice typing into a spoken instruction to rewrite the selection. Disable it to avoid selection-triggered rewrites. Bind direct voice typing when you want to skip polish and retain the recognized wording.
Models
Section titled “Models”Cloud options include OpenAI, Anthropic, NVIDIA NIM, DeepSeek, Volcengine Ark, Qwen, Kimi, GLM, MiniMax, MiMo, Grok, Gemini, OpenRouter, Inception, and OpenAI-compatible endpoints. Check the app for available choices. Cloud services usually need your own API key and receive the text being processed.
You can also download the Sharp-Spark-X2.5-4B local model and runtime from Text processing. Once prepared, inference runs on-device without an API key. Initial downloads need a network connection; speed and memory needs depend on hardware.
Learning & whole-utterance shortcuts
Section titled “Learning & whole-utterance shortcuts”Repeated stable polish results for the same whole utterance in ordinary voice typing can produce a learned rule. When a later complete transcript equals the rule’s original after normalization, the app outputs the rule’s result and skips polish. A phrase appearing inside a longer sentence does not trigger this shortcut.
The overlay’s ⚡️0ms means no LLM call ran, rather than zero recognition and insertion time. Inspect, edit, delete, or convert learned rules to manual under Word replacements. This shortcut does not run in direct voice typing or voice translation.
Failure & cancellation
Section titled “Failure & cancellation”If a model is unconfigured, a request fails, or Local Only blocks it, ordinary polish shows a reason and falls back to the recognized text for local processing and output. Local Only can use installed local models or allowed loopback endpoints.
Esc during polish attempts to deliver the original transcript when available. Cancelling a selection rewrite does not insert the spoken instruction. See Shortcuts & recording.
Voice translation uses a separate shortcut and can use the text-processing model or Apple native translation on macOS 26+.