An hour of audio transcribed in about 13s on MacBook Air
Summary
The article showcases EmbeddingGemma 2, a multimodal embedding model by Google DeepMind repurposed for on-device speech-to-text on a base MacBook Air, transcribing an hour of audio in roughly 13 seconds. It covers how the on-device reader maps audio slices to a word embedding space, benchmarks against other models, and notes a 5% word error rate with potential improvements, along with privacy benefits since audio stays on-device and the Apache 2.0 release. Implications include faster, offline transcription for SMB IT and potential shifts in on-device ASR tooling.