An hour of audio transcribed in about 13s on MacBook Air (eg2.haku.sh)

Google's EmbeddingGemma 2 has transformed an embedding model into a remarkably efficient speech-to-text system, achieving one hour of audio transcription in just about 13 seconds on a base MacBook Air. This breakthrough is significant for the AI/ML community as it not only demonstrates the versatility of embedding models—traditionally used for tasks like image and text processing—but also showcases how they can perform high-speed audio transcription entirely on-device, protecting user privacy by keeping data local. The process involves an audio encoder that remains unchanged and a small reader trained on under two hours of openly licensed English speech data. As short audio slices are processed, they are matched to the closest word from a collection of 229,000 English words, resulting in a transcript. This innovation achieves a word error rate of approximately 5%, which is in line with current leading models but suggests potential for further improvement with enhanced training. Users can easily transcribe audio or dictate into applications directly, marking a significant advancement in the accessibility and functionality of speech recognition technology on consumer devices.