The project began with Google’s EmbeddingGemma 2, which the source describes as an embedding model that handles text, images, audio and video in a shared vector space. Because audio clips of spoken words land near their written forms in this space, the team tested whether reading the nearest words directly from the embedding space could yield a transcript.
According to the source, the audio encoder of EmbeddingGemma 2 was left unchanged. A small reader network moves short slices of audio into the model’s text space, selects the closest of approximately 229,000 English words for each slice, and assembles those words into a raw transcript. A separate tiny model then adds punctuation and capitalization. Only the reader was trained; this took under two hours on eight H100 GPUs using openly licensed English speech. The base EmbeddingGemma 2 model itself never received any speech‑to‑text training.
Performance numbers are provided in a table. On a base M3 MacBook Air with 8 GB of RAM and the Neural Engine, the system processes one hour of audio in about 12.6 seconds, which the source expresses as 288× real‑time speed. For comparison, IBM Granite Speech 5.0 TurboCTC runs at roughly 150× real‑time (25.7 seconds per hour) on the same hardware. Published benchmarks for NVIDIA Parakeet TDT 0.6B v3 (FluidAudio) show 156× on an M4 Pro, 335× on an M5 Pro, and a fine‑tuned variant called Phonon‑2 reaches 480× on an M5 Pro.
The source reports the current word error rate of the system is about 5 %, while leading speech models achieve 2–3 %. The authors note that a better‑trained reader could narrow this gap and plan to pursue that in the next iteration.
For end users, the application works by dropping an audio file into the interface or by holding the Right Option key, speaking, and releasing to dictate into any macOS app. Because the computation runs on the Mac’s Neural Engine, the audio never leaves the device, and the model downloads (approximately 715 MB) on first launch.
All claims above are drawn directly from the provided source text; no external information or speculation has been added.
Key facts
- : The system transcribes one hour of audio in roughly 13 seconds on a base M3 MacBook Air.
- Trained reader only: under two hours on 8×H100, base model unchanged.
- Word error rate ~5 % versus 2–3 % for top speech‑to‑text models.
- Model size ~715 MB, runs locally on Neural Engine, audio stays on‑device.