GoKawiilDeveloper repurposes Google's EmbeddingGemma 2 into offline speech-to-text Mac app
A developer has released a free macOS app that converts Google's EmbeddingGemma 2 multimodal embedding model into a speech-to-text tool, transcribing an hour of audio in about 13 seconds on a base MacBook Air. The system keeps the original audio encoder untouched and instead trains a small 'reader' model to map audio slices into the text embedding space, matching them to the closest of roughly 229,000 English words, with a separate model adding punctuation. The app, EmbeddingGemma Transcribe, runs entirely on-device using the Mac's Neural Engine and currently has a word error rate of about 5%, compared with 2-3% for leading dedicated speech models.
GoKawiil's interpretation of the reporting above, not reported fact.
The project suggests embedding models, originally designed to map different data types like text and audio into a shared space, could be repurposed for tasks like transcription without retraining the core model, potentially lowering the cost of building niche speech tools. Running entirely on-device also means no audio leaves the machine, which could appeal to privacy-conscious users, though the current accuracy gap versus specialized models may limit practical use until the developer's planned improvements to the reader model are completed.
- EmbeddingGemma Transcribe claims to transcribe an hour of audio in about 13 seconds on a base M3 MacBook Air using only the Neural Engine.
- Only a small 'reader' component was trained, in under two hours on 8×H100 GPUs, while Google's original embedding model was left unchanged.
- Its word error rate is around 5%, higher than the 2-3% typical of leading speech-to-text models, with the developer saying further training could narrow the gap.
Apple MacBook Air M3 — This article's blazing-fast on-device transcription runs natively on Apple Silicon, and the base M3 MacBook Air is exactly the machine benchmarked here. Its efficient Neural Engine lets you run cutting-edge local AI models like EmbeddingGemma without needing a cloud connection or heavy GPU rig. It's a great pick if you want to experiment with on-device speech-to-text tools yourself.
See Apple MacBook Air M3 on Amazon → Affiliate link — we may earn a commission on purchases, at no extra cost to you. Product picked by AI based on this article; it is not a tested recommendation.Source: eg2.haku.sh, 2026-10-13
Published there as: “An hour of audio transcribed in about 13s on MacBook Air”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.