Player not loading? Watch on YouTube
FluidVoice is presented as an open source macOS dictation app that processes speech offline. The walkthrough explains its architecture rather than providing a step-by-step installation guide. The speaker says audio and text stay on the user's hardware and describes a zero-telemetry environment.
The stated requirements are Apple Silicon M1 through M4 and macOS 15.0 or later. Acoustic models remain loaded in memory to reduce startup delays. According to the speaker, a full local installation occupies nearly 5 GB of active RAM; a compressed 1 GB mini model for Macs with 8 GB of memory is described as under development.
The pipeline separates speech recognition from text cleanup. Core ML acoustic models produce raw text, then Fluid-1, a proprietary local LLM, handles punctuation, hesitation markers, and corrections. The speaker describes Fluid-1 as a 3.5 GB model trained on 100,000 dictation examples. A cited benchmark favors it over GPT 5.4, but the transcript does not explain the evaluation method.
User-defined vocabulary replacements run before AI processing. The walkthrough describes a JSONC mapping file that reloads after edits without a daemon restart. FluidVoice then uses macOS Accessibility access to type into the active window. It also discusses microphone permission failures and a terminal-based reset, though the transcript supplies no command to reproduce that fix.
