EmbeddingGemma 2 converted to 13s-per-hour speech transcription tool
October 13, 2026
By training a small reader to map audio slices to EmbeddingGemma 2's text space, researchers achieved 1-hour audio transcription in 12.6 seconds on an M3 MacBook Air. This on-device approach utilizes a shared multimodal embedding space rather than traditional end-to-end training.
HOW THIS AFFECTS YOU
●
builderYou can deploy extremely fast, low-latency, on-device transcription using embedding models.
●
researcherThis demonstrates the efficiency of using cross-modal embedding spaces for speech tasks.