SantiSotto
Privacy-first voice transcription for macOS
At a Glance
Built for Speed and Privacy
*With tiny models on Apple Silicon
How It Works
Three Steps. Zero Cloud.
Hold your hotkey to record, release to transcribe. Text appears at your cursor instantly. The entire pipeline runs on your Mac.
Hold Hotkey
Press Option+Space to start recording from your microphone
Capture Audio
Audio captured at 48kHz, resampled to 16kHz in real-time
Whisper AI
Local Whisper model transcribes using Metal GPU acceleration
Post-Process
Remove fillers, clean punctuation, apply dictionary bias
Paste Text
Text inserted at cursor via clipboard, original clipboard restored
Features
Everything You Need
Designed to be invisible until you need it. Runs in your menu bar, stays out of your way.
Complete Privacy
All audio processing happens on-device. No internet required for transcription. No data collection, no telemetry, no cloud APIs.
Metal GPU Acceleration
Leverages Apple Silicon's Metal API for hardware-accelerated inference. Transcription in under a second with small models.
30 Whisper Models
From 31MB tiny to 2.9GB large-v3. Full precision and quantized variants (Q5, Q8). English-only models for extra speed.
Transcription History
Every transcription saved locally. Browse, copy, or delete individual entries. Never lose text if a paste fails.
Custom Hotkeys
Change your recording shortcut to any modifier+key combo. Capture UI with live key detection. Two configurable shortcuts.
Filler Word Removal
Automatically strips "um", "uh", "er", "you know", "basically" and more. Smart pattern matching avoids false positives.
Personal Dictionary
Add names, technical terms, and jargon that Whisper might not recognize. Biases the model toward your vocabulary.
99 Languages
Multilingual models auto-detect language. Supports English, Spanish, French, German, Japanese, Arabic, and 93 more.
Menu Bar App
Runs as a macOS accessory — no dock icon, no window clutter. Launches at login. Always ready when you are.
AI Models
Choose Your Trade-off
Larger models are more accurate but slower. Quantized variants (Q5, Q8) reduce size with minimal quality loss. English-only models are faster for English.
Tiny
31 – 75 MB
Base
57 – 142 MB
Small
181 – 466 MB
Medium
514 MB – 1.5 GB
Large v3
547 MB – 2.9 GB
Why SantiSotto
Compared to Alternatives
Most transcription tools send your audio to the cloud. SantiSotto keeps everything on your device.
| Feature | SantiSotto | Otter.ai | macOS Dictation | Whisper (CLI) |
|---|---|---|---|---|
| 100% Local Processing | ✓ Yes | ✗ Cloud | ● Hybrid | ✓ Yes |
| No Account Required | ✓ Yes | ✗ Required | ✓ Yes | ✓ Yes |
| Works Offline | ✓ Yes | ✗ No | ● Limited | ✓ Yes |
| GPU Accelerated | ✓ Metal | N/A (cloud) | ✓ Yes | ● Manual setup |
| Hotkey Recording | ✓ Customizable | ✗ No | ● Fixed key | ✗ No |
| Auto-Paste at Cursor | ✓ Yes | ✗ Copy only | ✓ Yes | ✗ File output |
| Model Selection | ✓ 30 models | N/A | ✗ Fixed | ✓ Manual |
| Transcription History | ✓ Local | ✓ Yes | ✗ No | ✗ No |
| Filler Word Removal | ✓ Automatic | ● Basic | ✗ No | ✗ No |
| Free | ✓ Open source | ✗ Subscription | ✓ Yes | ✓ Open source |
| GUI / Easy Setup | ✓ Native app | ✓ Web app | ✓ Built-in | ✗ Terminal only |
Under the Hood
Technology Stack
A modern, performant architecture built with native technologies.
Frontend
TypeScript + Vite for fast, type-safe UI development. Lucide icon library for consistent, beautiful iconography. Pure CSS with light and dark mode support.
Backend
Rust for memory-safe, high-performance audio processing and transcription. Tauri v2 framework for native macOS integration with minimal overhead (~5MB binary).
AI Engine
whisper-rs (whisper.cpp bindings) with Metal feature flag for Apple Silicon GPU acceleration. Greedy sampling for lowest latency. Auto language detection for multilingual models.
Audio Pipeline
cpal for cross-platform audio capture. Real-time 48kHz to 16kHz downsampling via sinc-based resampling with rubato. Zero-copy buffer management with privacy-first memory clearing.