Embedding Gemma Transcribe hits 288× realtime

This is a clever local-inference experiment, trading some accuracy for extraordinary speed, privacy, and low hardware requirements.

• –Reuses EmbeddingGemma 2’s audio encoder and trains a small reader to map speech slices into text space, selecting from roughly 229,000 English words.
• –The 288× real-time result on an 8GB MacBook Air makes batch transcription practical without cloud APIs, discrete GPUs, or recurring costs.
• –Its roughly 5% word error rate still trails mature ASR systems, limiting it for noisy audio, multiple speakers, and high-stakes transcription.
• –The project demonstrates a useful developer pattern: adapt multimodal embedding backbones to narrow on-device tasks with lightweight task-specific heads.
• –Google’s underlying EmbeddingGemma 2 is a 740M-parameter Apache 2.0 model designed for unified text, image, audio, and video embeddings. [Google announcement](https://blog.google/innovation-and-ai/technology/developers-tools/embeddinggemma-2/?authuser=399)