Speak naturally. Let offline AI models clean, format, and structure your speech in real-time. Instantly pasted at your cursor. Zero subscriptions. Zero cloud tracking. Just pure speed.
We built this because paying $12/month to a venture-backed cloud startup to type your own dictations on your own machine is a crime.
No network calls, no cloud telemetry, no sound recordings uploaded to remote servers. Everything runs entirely offline on your SSD.
Transcribe and clean raw speech with native Metal GPU acceleration via Apple MLX, completing in under 300 milliseconds.
Whisper + LLM mapping captures specific Indian names, stutters, Hinglish grammar structures, and jargon without generic rephrasing.
Every single dictation's high-fidelity audio is saved locally. Play back, scrub, and copy transcripts inside your native dashboard.
Using native AppKit modifier monitoring, trigger dictation instantly by holding Control without triggering macOS accessibility loops.
Runs entirely in the background as an agent. Hides from your Dock and shows clean SF Symbols menu bar states.
See how phisphis takes raw, stutter-filled Hinglish speech, cleans out filler words, and turns it into clean formatted prose instantly.
phisphis is local, lightweight, and takes just a couple commands to launch.
Install Ollama on your Mac and pull the lightweight models for formatting:
ollama pull qwen2.5:1.5bGrab the codebase to get setup with python environment launcher:
git clone https://github.com/ChawlaBhai/phisphis.git
cd phisphis
./run.shPackage it into a native macOS status bar application bundle:
.venv/bin/python -m PyInstaller phisphis.spec --clean -y
# Bypass unidentified developer block
xattr -d com.apple.quarantine dist/phisphis.app
open dist/phisphis.appWe are actively building the future of local, zero-cost, hyper-private productivity engines.
Training an ultra-compact speech-to-text model specifically compiled to run inside Apple Neural Engine (ANE) cores for blazing-fast 0-latency execution.
Teach the formatting AI your unique team slang, code symbols, variables, and project structures for custom auto-formatting.
Deploying the exact same zero-cloud, hardware-accelerated workflow on Windows (DirectML) and Linux (CUDA/ROCm).