Embedding Gemma Transcribe

Embedding Gemma Transcribe (EG2 Transcribe) is a free, native macOS 14+ app for Apple Silicon that turns speech into text entirely on device — the audio never leaves the machine. It runs on the Neural Engine and, per its maker, transcribes an hour of audio in about 13 seconds on a base MacBook Air.

It is built on EmbeddingGemma 2, a multimodal embedding model from Google DeepMind released under Apache 2.0, which was never trained for speech recognition. The audio encoder is left unchanged and only a small "reader" was trained to move short audio slices into the model's text space; the closest of about 229,000 English words forms the transcript, and a tiny extra model adds punctuation and capitalisation. Its word error rate is currently around 5%.

The app accepts dropped audio files and push-to-talk dictation (hold Right Option) into any application. The models (~715 MB) download on first launch. It is made by Haku and is an independent project, not affiliated with or endorsed by Google.