Today, we are thrilled to announce the release of Escribolt 2.0, a major evolution in how you capture, process, and organize your spoken thoughts natively on macOS. With version 2.0, we have integrated local Whisper pipelines and local reasoning models directly into the application—offering true, offline-first voice intelligence.
The Case for Local-First AI
In an era dominated by API-based cloud AI wrappers, user privacy and speed can often take a back seat. When you dictate a confidential meeting summary or record a proprietary design sync, you shouldn't have to trust that remote servers won't leak your transcripts or train models on your internal data.
Escribolt 2.0 treats privacy as an architectural fact, not just a checkbox. In Local Mode, 100% of speech processing and reasoning occurs directly on your physical Mac. Your audio inputs are written to local buffers, transcribed on your processor, and saved directly into a secure local SQLite database.
Under the Hood: Our On-Device Stack
To make local intelligence practical on standard consumer computers, we designed an ultra-efficient inference layer:
- Speech-to-Text: Powered by a native Whisper.cpp implementation. It is optimized to leverage the Apple Neural Engine (ANE) on M-series chips, delivering transcription speeds up to 4x faster than real-time.
- Language Model (LLM): Runs local weights for Qwen-2.5-Instruct (or Google Gemma) natively. We quantize the models down to 4-bit, maintaining strong reasoning capabilities while consuming less than 1.5GB of system RAM.
How to Get Started
Upgrading to version 2.0 is simple. When you launch the app, select Local Mode during the onboarding flow. Escribolt will run a one-time bootstrap script to fetch the model weights (about 1.8GB) and verify your hardware compatibility.
Once initialized, you can turn off your Wi-Fi, toggle your hold-to-talk hotkey, and speak with complete peace of mind—even on an airplane.