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

Google has created a speech-to-text tool using their EmbeddingGemma 2 model that transcribes one hour of audio in approximately 13 seconds on a MacBook Air, running entirely on-device without sending data to external servers. The model works by treating audio as embeddings in a shared space with text, matching audio slices to the closest English words from a vocabulary of ~229,000 words, then adding punctuation through a separate small model. While it currently has a 5% word error rate compared to 2-3% for leading speech models, the developers believe this can be improved with better training in future iterations.