tl;dr: Developers repurposed Google's EmbeddingGemma 2 multimodal embedding model into an on-device speech-to-text system by training a small "reader" that maps audio slices to the nearest of ~229,000 English words in the shared embedding space. On a base M3 MacBook Air it transcribes an hour of audio in 12.6 seconds (288× real-time) using the Neural Engine, roughly twice as fast as IBM Granite Speech 5.0 Turbo. Word error rate is around 5%, versus 2–3% for leading STT models, which they hope to close with better reader training.