Embedding Gemma Transcribe

About

A free Mac app that turns Google's EmbeddingGemma 2 embedding model into a speech-to-text system. A small trained reader maps audio slices into the model's text space and picks the nearest English word, with a separate model adding punctuation. The developers report about an hour of audio in 13 seconds on a base MacBook Air and about 5% word error rate, higher than the 2 to 3% of leading speech models.