Engineer repurposes EmbeddingGemma 2 for speech recognition

On October 12, 2026, engineer Shahir, who works on machine learning directly on devices, said on X that he had repurposed Google's EmbeddingGemma 2 embedding model for speech recognition. According to him, the result is one of the fastest STT models that can be run on-device.

The idea appeared in March together with Gemini Embedding 2: it places text, images, audio and video into one shared space, where a spoken word ends up next to a written one. He could not try it back then, but last week EmbeddingGemma 2 was released as an open model, and the author tested the hypothesis.

The model itself was not modified. A small "reader" works on top of it: it takes the next fragment of audio, finds the nearest of roughly 229,000 words in the model's own text space, and assembles them into text.

According to Shahir, on a base MacBook Air with an M3 chip, an hour of audio is transcribed in about 13 seconds, that is, 288× real time. Training took less than two hours on 8×H100, and openly licensed speech was used for it.

So that the method could be tested, the author released a free Mac app.

Translation editor: Салтанов А.