Researchers have repurposed Google's EmbeddingGemma 2 model to create a speech-to-text transcription tool that runs entirely on a MacBook Air. This on-device system can transcribe an hour of audio in approximately 13 seconds, though its current word error rate is around 5%, compared to 2-3% for leading models. The team trained a small "reader" component to map audio slices into the model's existing text embedding space, achieving this functionality without retraining the core EmbeddingGemma 2 model. AI
IMPACT Demonstrates novel applications for multimodal embedding models and enables efficient on-device transcription.
RANK_REASON Repurposing of an existing embedding model for a new AI task (speech-to-text) with performance benchmarks. [lever_c_demoted from research: ic=1 ai=1.0]
Read on Hacker News — AI stories ≥50 points →
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
