WhisperForge

Fast, private dictation for macOS — speak in any language and get polished technical English pasted wherever you type.

On-device speech recognition (Parakeet v3 / Whisper) plus an optional LLM pass that rewrites your rough dictation into sharp, professional English. Local or cloud.

A free, open-source SuperWhisper alternative for macOS — no subscription, runs fully on-device, with an optional AI rewrite that turns rough multilingual speech into clean technical English. A privacy-friendly alternative to SuperWhisper, Wispr Flow, and MacWhisper.

A fork of OpenSuperWhisper, adding an AI post-processing pipeline.

Highlights

• 🎙️ Real-time dictation with a global hotkey (hold-to-record)
• 🧠 On-device ASR — Parakeet v3 (multilingual, ANE-accelerated) or Whisper
• 🤖 AI post-processing — rewrite/translate the transcript before pasting (e.g. rough Portuguese → polished technical English)
• ☁️ Pluggable LLM — local (Ollama / LM Studio) or cloud (Groq, OpenAI, OpenRouter, Anthropic) via an OpenAI-compatible base URL
• 🔒 Private by default — API keys in the Keychain; dictation audio is discarded right after transcription (no history pile-up)
• 📋 Auto-paste into the focused app; drag-and-drop file transcription

How it works

speak → on-device ASR transcribes → (optional) LLM rewrites to polished English → pasted

History is off by default: the audio is deleted the moment it is transcribed, so nothing accumulates on disk.

AI post-processing setup

In Settings → AI Post-processing:

1. Toggle Enhance with AI.
2. Pick a Provider and paste an API key (stored in the Keychain, never on disk).
3. Leave Model, Base URL, and System Prompt empty to use sensible defaults.

Recommended on low-RAM Macs: cloud Groq (free tier) with openai/gpt-oss-120b — near-zero local RAM and no model-load spike. A local LLM can exhaust a 16 GB Mac during dictation, so prefer cloud on constrained hardware; the local ASR (Parakeet) stays on-device and is light.

The built-in system prompt rewrites rough multilingual dictation into sharp, senior-engineer technical English — adapting the form to the content (commit, message, or spec) and dropping phrases the speaker retracts.

Build from source

Requires Xcode 26+ and Homebrew packages cmake, libomp, rust.

git clone <this-repo-url>
cd OpenWhisper
git submodule update --init --recursive
brew install cmake libomp rust
./run.sh build        # builds the app; `./run.sh` (no arg) also launches it

run.sh builds the whisper.cpp and Rust autocorrect dependencies, then the app. It also applies an idempotent patch so the pinned FluidAudio (0.11.0) compiles under Xcode 26 / Swift 6 (see Scripts/patch_fluidaudio.py).

Requirements

• macOS 14+ (Apple Silicon / ARM64)

Documentation

• docs/LLM_FEATURE_REVIEW.md — design, code review, and rationale for the AI post-processing pipeline.

Credits

This fork (WhisperForge) is built and maintained by @vkmtx.

Built on OpenSuperWhisper by Starmel. Speech models via whisper.cpp and FluidAudio. Asian autocorrect via autocorrect.

License

MIT — inherited from the upstream project. See LICENSE.

---

_{Keywords: macOS dictation · voice to text · speech to text · open-source SuperWhisper alternative · free SuperWhisper alternative · Wispr Flow alternative · MacWhisper alternative · local Whisper dictation · Parakeet · on-device transcription · AI dictation · privacy-first speech-to-text for Mac.}