Embeddings become speech-to-text

An unconventional project has turned Google’s EmbeddingGemma 2, a model built to place text, images, audio and video in a shared space, into a speech-to-text tool. The base model wasn’t trained for transcription. Instead, a small reader maps short audio slices into the model’s text space, then selects the closest match from about 229,000 English words. A small extra model adds punctuation and capitals. The project says only the reader was trained, and the base model wasn’t trained for speech-to-text.

The result runs on a Mac’s Neural Engine, with audio staying on the device. The project reports transcribing an hour of audio in 12.6 seconds on a base M3 MacBook Air, though that speed isn’t the whole story. Its current word error rate is about 5%, compared with 2–3% for leading speech models, and the project says improving the reader is the next step. So this is an interesting option to watch, not a reason to swap out a transcription system without testing it on your own audio. The project page has the details, and its Hacker News discussion drew 337 points and 79 comments. This is the kind of development we build into an AI command centre, so a business can track practical tools while checking their limits.