EmbeddingGemma 2 repurposed as speech-to-text
Original: We turned EmbeddingGemma 2 into a speech-to-text model
Why This Matters
Shows embedding models hold untapped capability beyond their intended use cases.
Independent developers converted Google DeepMind's EmbeddingGemma 2 embedding model into an on-device speech-to-text tool, transcribing one hour of audio in 12.6 seconds on a base M3 MacBook Air with 8 GB RAM.
EmbeddingGemma 2 was never built for transcription — it's a multimodal embedding model that maps text, images, audio, and video into a shared vector space. A small independent team exploited that architecture: spoken words land geometrically close to their written equivalents in the embedding space, so they trained a lightweight 'reader' to pull transcriptions directly from it. The base model is untouched. Only the reader was trained, taking under two hours on 8×H100 GPUs using openly licensed English speech data. A second tiny model handles punctuation and capitalization. Speed is the headline. On a base M3 MacBook Air, it hits 288× real-time — one hour of audio in 12.6 seconds — running entirely on Apple's Neural Engine. IBM Granite Speech 5.0 Turbo manages about 150× on the same machine. The tradeoff is accuracy: word error rate sits at ~5%, versus 2–3% for leading speech models. The team says a better-trained reader should close that gap. The Mac app is free, supports file drop and live dictation via Right Option key, and keeps all audio on-device. EmbeddingGemma 2 is Apache 2.0 licensed; this project is independent of Google.