An hour of audio transcribed in about 13s on MacBook Air

This describes a Mac app that turns Google's multimodal embedding model EmbeddingGemma 2 into an on-device speech-to-text system that can transcribe an hour of audio in roughly 12-13 seconds on a base M3 MacBook Air using the Neural Engine. The approach leaves EmbeddingGemma’s audio encoder untouched and trains a small reader that maps short audio slices into the model’s shared text embedding space; each slice is matched to the nearest word from a vocabulary of about 229,000 terms, and a tiny secondary model applies punctuation and capitalization. Only the reader was trained (under two hours on 8×H100 GPUs using openly licensed English speech); the base embedding model was never trained for speech recognition. The app requires macOS 14+ on Apple Silicon, downloads a ~715 MB model at first launch, and runs fully locally.

Measured performance shows extreme speed: the developer reports 288× real time (one hour transcribed in ~12.6 s) on a base M3 MacBook Air, compared to about 150× (25.7 s) for an IBM Granite Speech run on the same machine; other published FluidAudio benchmarks on newer Macs show different multipliers. Word error rate is around 5%, versus roughly 2-3% for top speech models, and the team expects improvements with a better-trained reader. The software is free, preserves privacy by keeping audio on device, and is an independent project using EmbeddingGemma 2 (Apache 2.0) from Google DeepMind.