An independent experiment has adapted Google’s EmbeddingGemma 2 embedding model into a speech-to-text system without retraining the base model. Its creators report that their prototype can process an hour of audio in roughly 13 seconds on a base M3 MacBook Air while keeping the recording on the device.
EmbeddingGemma 2 was built to place text, images, audio and video into a shared mathematical space. Similar concepts across those formats should end up near one another: the sound of a spoken word, for example, can be represented close to the written word. The project, called Embedding Gemma Transcribe, tests whether a lightweight reader can recover text directly from that shared representation.
Only the reader was trained. The developers say training took less than two hours on eight Nvidia H100 accelerators and used openly licensed English speech. The underlying EmbeddingGemma 2 model was left unchanged and was not originally trained as a speech-recognition model. That makes the project a demonstration of how general-purpose multimodal representations can be repurposed, rather than a conventional end-to-end transcription model.
The team reports a word error rate of about 5 percent. It says leading dedicated speech models reach roughly 2 to 3 percent, so the prototype does not yet match the best systems on accuracy. The developers believe a better-trained reader could narrow that gap, but that remains a target for future work rather than a demonstrated result.
Performance figures also need careful interpretation. The project compares its own M3 MacBook Air tests with published FluidAudio benchmarks collected on larger, newer Macs and different audio. Its website explicitly describes that table as a rough guide, not a like-for-like ranking. The directly supported claim is the team’s own result: approximately 13 seconds to process one hour of audio on its base M3 machine.
The application offers file transcription and a push-to-talk dictation mode. A user can hold the Right Option key, speak and release it to send recognized text into another application. Inference runs on the Mac’s Neural Engine, and the developers say audio does not leave the computer. Those characteristics make privacy and latency, rather than benchmark-leading accuracy, the clearest attractions of the prototype.
The experiment suggests that a multimodal embedding model may contain enough alignment between speech and text for a relatively small decoder to extract useful transcripts. It does not show that any embedding model can be converted this way, nor does it replace controlled evaluation across languages, accents and noisy environments. For now, it is a focused proof of concept with unusually high reported speed, a disclosed accuracy gap and a clear path for further training.
