Google’s EmbeddingGemma 2, a multimodal embedding model designed to work with text, images, audio, and video in a shared space, has been adapted into a speech-to-text system capable of transcribing an hour of audio in approximately 13 seconds on a base MacBook Air with an M3 chip and 8 GB of RAM.
The approach repurposes the model’s existing audio encoder without retraining it. Instead, a small reader component maps short audio slices into the model’s text space, where they are matched against roughly 229,000 English words. A separate small model then adds punctuation and capitalization. Only the reader was trained, using less than two hours of processing on 8×H100 GPUs with openly licensed English speech data.
According to the source, Embedding Gemma Transcribe achieves 288 times real-time speed—meaning it processes 288 seconds of audio per second of computation—on a base M3 MacBook Air. This outpaces IBM Granite Speech 5.0 TurboCTC, which runs at approximately 150 times real-time on the same hardware. The model downloads approximately 715 MB on first launch.
The transcription system operates entirely on-device using the Mac’s Neural Engine, ensuring audio never leaves the machine. Users can either drop in an audio file or use a keyboard shortcut (Right Option held while speaking) to dictate into any application.
According to the source, the current word error rate stands at approximately 5%, compared to 2–3% for leading speech-to-text models. The developers indicate this gap can be closed with a better-trained reader in future iterations. Benchmarks from other models on newer or different Mac hardware are cited as rough guides rather than direct comparisons, since testing conditions and hardware vary.
Key facts
- Embedding Gemma Transcribe processes an hour of audio in 12.6 seconds on a base M3 MacBook Air
- The system uses only a trained reader component; the base embedding model was never trained for speech-to-text
- All processing occurs on-device via the Mac’s Neural Engine
- Current word error rate is approximately 5%
- The model downloads 715 MB and only the reader required training on 8×H100 GPUs