Open Source
Explore the latest AI open-source projects from GitHub and HuggingFace.
Explore the latest AI open-source projects from GitHub and HuggingFace.
TypeWhisper for Mac is an open-source, on-device speech-to-text and AI text-processing app for macOS that turns any text field on the system — Mail, Slack, Xcode, a browser address bar — into a dictation target with Apple Silicon-optimized Whisper, Parakeet, and Apple SpeechAnalyzer models behind it. Released under GPL-3.0 and distributed as a universal binary, the project has reached 1,216 stars and 81 forks and ships through both the Homebrew tap `typewhisper/tap/typewhisper` and direct DMG downloads. ## The macOS Dictation Gap Apple's built-in dictation has improved in recent macOS releases, but it remains tightly tied to Siri, lacks workflow hooks, and offers no swap-in support for state-of-the-art ASR models like Distil-Whisper or Parakeet TDT v3. Power users — developers, writers, accessibility-dependent users, and anyone whose hands hurt by 3pm — have for years relied on third-party tools like Aiko, MacWhisper, and Superwhisper to fill the gap. TypeWhisper enters that space with an explicit "local first, cloud optional" stance and a plugin architecture aimed at automation. ## Eleven Pluggable Transcription Engines The headline feature is the breadth of supported back-ends. WhisperKit, Argmax's Apple-Silicon-optimized port of OpenAI Whisper, handles 99+ languages with streaming output. NVIDIA's Parakeet TDT v3 is included for 25 European languages with very low word-error rates. On macOS 26 and later, the app can dispatch to Apple's new on-device SpeechAnalyzer framework, which Apple ships with system-level acceleration. For users who need raw speed or specialized vocabulary, Groq's hosted Whisper, OpenAI's transcription API, and xAI/Grok's audio endpoint can be plugged in with API keys — and TypeWhisper keeps every key in the macOS Keychain rather than a plaintext config file. ## System-Wide Dictation and Per-App Workflows A global hotkey activates dictation anywhere a text caret can sit. What makes TypeWhisper distinctive is the workflow layer wrapped around that primitive. Workflows are reusable transformations — "transcribe, then summarize as bullet points," "transcribe, then translate to German," "transcribe, then format as Conventional Commit message" — and they can be bound to specific apps or even specific websites via per-domain triggers. Dictate into a Slack window and a friendly, casual workflow runs; dictate into Xcode and a technical, precise one takes over. ## Local LLM Post-Processing with Gemma 4 For users who want AI cleanup of transcribed text without sending it anywhere, TypeWhisper bundles support for Google's Gemma 4 family running locally through MLX. Dictated text can be auto-punctuated, de-ummed, formatted, or rewritten in a target tone, all on-device. The same pipeline supports Ollama as an alternative LLM backend, giving users the choice of Llama, Qwen, or Mistral models for the post-processing step. ## Plugin Registry and Automation Hooks The app exposes an HTTP API on localhost and a CLI tool so other applications — Raycast, Alfred, Keyboard Maestro, a shell script — can drive dictation programmatically. A community plugin registry lets users install third-party workflows with one click, and a dictionary plus snippets system handles names, acronyms, and proper-noun corrections that confuse Whisper-style models out of the box. ## Pros, Cons, and Audience The app is unmistakably Mac-native: SwiftUI front-end, universal binary, Sonoma+ requirement, and 8GB RAM minimum with 16GB recommended for the larger models. There is no Windows or Linux build, and the GPL-3.0 license may discourage proprietary enterprise integration. UI localization is limited to English and German for now, even though transcription itself spans 99+ languages. For Mac power users who want a free, scriptable alternative to MacWhisper Pro or Superwhisper — with the option to keep everything on-device — TypeWhisper is one of the most actively maintained options in the 1k-star tier on GitHub.