An embedding model transcribes an hour of audio in 13 seconds on a base MacBook Air

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

Google's EmbeddingGemma 2, a multimodal embedding model, was repurposed for speech-to-text. A small reader maps audio slices into the model's text space, matching against ~229,000 English words, with a tiny model adding punctuation. The result runs entirely on-device on a base M3 MacBook Air, transcribing an hour of audio in 12.6 seconds—faster than IBM Granite Speech and NVIDIA Parakeet on similar hardware. Word error rate is about 5%, with plans to improve.