团队把 EmbeddingGemma 2 改造成端侧语音转写工具并发布 Mac 应用 v1.6

Google's EmbeddingGemma 2 is an embedding model. We turned it into a speech-to-text model: an hour of audio in about 13 seconds on a base MacBook Air, entirely on device.

Download for Mac (free) v1.6 · macOS 14+ · Apple Silicon
models download on first launch (715 MB)

Can an embedding model transcribe speech?

EmbeddingGemma 2 is a multimodal embedding model: text, images, audio and video in one shared space. Say a word out loud and its audio lands close to the written word.

So we tried reading words straight out of that space.

How it works

EmbeddingGemma 2's audio encoder stays exactly as it is.
A small reader moves each short slice of audio into the model's text space.
The closest of ~229,000 English words wins, and the slices become a transcript.
A tiny extra model adds punctuation and capitals.

Only the reader is trained: under two hours on 8×H100, on openly licensed English speech. The base model was never trained for speech-to-text.

How fast?

Model	Runs on	Mac	Speed
Embedding Gemma Transcribe	Neural Engine	base M3 MacBook Air, 8 GB	288×1 hour in 12.6 s
IBM Granite Speech 5.0 TurboCTC	GPU (MLX)	base M3 MacBook Air, 8 GB	~150×1 hour in 25.7 s
NVIDIA Parakeet TDT 0.6B v3 (FluidAudio)	Neural Engine	M4 Pro	156×
NVIDIA Parakeet TDT 0.6B v3 (FluidAudio)	Neural Engine	M5 Pro	335×
Phonon-2, a Parakeet v3 fine-tune (FluidAudio)	Neural Engine	M5 Pro	480×

× real time: seconds of audio per second of processing. Rows 1–2: our tests on the same base M3 MacBook Air. Rows 3–5: FluidAudio's published benchmarks on larger, newer Macs and different audio. A rough guide, not a like-for-like ranking.

Where it stands

Its word error rate is about 5% right now, against 2–3% for leading speech models. We believe a better-trained reader can close that gap; that's the next iteration.

Use it

Drop in an audio file, or hold Right Option, speak and release to dictate into any app. It runs on the Mac's Neural Engine; your audio never leaves the machine.