Embedding model prototype claims hour-long transcription in about 13 seconds

The developers of Embedding Gemma Transcribe say their speech-to-text prototype processes an hour of audio in about 13 seconds on a base MacBook Air, entirely on device. It keeps EmbeddingGemma 2's audio encoder unchanged and trains a small reader to map short audio slices into the model's text space, selecting the closest English word. Only the reader was trained, using openly licensed English speech in under two hours on 8×H100. A separate tiny model adds punctuation and capitals. The developers report a word error rate of about 5%, against 2-3% for leading speech models, and caution that external speed benchmarks used different hardware and audio.