# WisprType — Voice to text, anywhere on macOS

> A native macOS dictation app powered by on-device AI. Hold a key, speak,
> and your words appear wherever you type. Private by default, blazingly
> fast on Apple Silicon.

- **Website:** https://wisprtype.com
- **Platform:** macOS 12+ (Apple Silicon)
- **License/Pricing:** Free during early access
- **Download:** https://wisprtype.com/WisprType_1.0.2_aarch64.dmg
- **Privacy policy:** https://wisprtype.com/privacy-policy
- **Contact:** contact@wisprtype.com — https://wisprtype.com/contact

## How it works

1. Hold the configured push-to-talk hotkey (Right ⌘ by default).
2. Speak naturally. Audio is captured locally.
3. Speech-to-text runs on-device via WhisperKit (Apple Silicon, Core ML).
4. Optional Smart Typing post-processing removes filler words, fixes
   self-corrections, and applies punctuation.
5. The cleaned transcript is pasted into the focused application.

## Engines

WisprType supports both local and cloud speech-to-text engines.

### Local (default, private)

| Model                   | Approx. size | Notes                  |
| ----------------------- | ------------ | ---------------------- |
| Whisper Tiny            | ~75 MB       | Fastest, lightest      |
| Whisper Base            | ~150 MB      | Default — recommended  |
| Whisper Small           | ~500 MB      | Better accuracy        |
| Whisper Medium          | ~1.5 GB      | High accuracy          |
| Whisper Large v3        | ~3 GB        | Best accuracy          |
| Distil-Whisper Large v3 | —            | Distilled variant      |

### Cloud (optional, opt-in)

| Provider | Default STT model        |
| -------- | ------------------------ |
| OpenAI   | gpt-4o-transcribe        |
| Groq     | whisper-large-v3-turbo   |
| Deepgram | nova-3                   |

Cloud providers are never used unless you explicitly configure an API key
in Settings → Models. Keys are stored locally and only used to authenticate
requests you initiate.

## Smart Typing

AI-powered transcript cleanup that removes filler words, fixes
self-corrections, and improves punctuation/formatting.

- **Local mode (default):** Runs Llama 3.2 3B locally via Apple's MLX
  framework. No transcript data leaves your machine.
- **Cloud mode:** Reuses your configured cloud provider for cleanup. Raw
  transcript text and dictionary words are sent to that provider only.

## Custom dictionary

Add domain-specific words across categories — `Name`, `Technical`,
`Medical`, `Legal`, `Other` — to bias recognition. The dictionary is
stored locally and never transmitted.

## Privacy summary

- Speech-to-text runs entirely on-device by default.
- No audio, transcripts, or personal content are sent to WisprType
  servers.
- Anonymous, opt-in usage telemetry only (PostHog, no IP, no transcript
  text).
- Cloud providers are opt-in and use API keys you supply.
- See the full [Privacy Policy](https://wisprtype.com/privacy-policy).

## System permissions

- **Microphone** — to capture audio for transcription.
- **Accessibility** — for the global hotkey and paste automation.
- **Audio Input** — entitlement for audio device access.
- **Apple Events** — to paste text into the active application.

## Discovery for agents

- API catalog (RFC 9727): https://wisprtype.com/.well-known/api-catalog
- Sitemap: https://wisprtype.com/sitemap.xml
- Web app manifest: https://wisprtype.com/manifest.webmanifest
- Agent skills index: https://wisprtype.com/.well-known/agent-skills/index.json
- Markdown via content negotiation: send `Accept: text/markdown` to any
  HTML route on this site.