EG2: An Embedding Model That Transcribes Speech at 288x Real Time

⬅️ Back to Tools

EG2 Embedding Gemma Transcribe

What it isspeech-to-text Mac app built from an embedding model, not a transcriber
PlatformmacOS 14+ on Apple Silicon
Pricefree
Linkeg2.haku.sh

Embedding models map words, images, and audio into one shared space, so spoken audio lands near its written word. Haku’s EG2 takes that literally: instead of training a transcriber, it reads words straight out of Google’s EmbeddingGemma 2 space. I downloaded the 2.5 MB disk image, mounted it, and confirmed version 1.6 sitting next to an Apache license and a Gemma use policy before writing this.

  1. The trick is the whole product. EmbeddingGemma 2’s audio encoder stays exactly as shipped. A small trained reader moves each short audio slice into the model’s text space, the closest of roughly 229,000 English words wins, and a tiny extra model adds punctuation and capitals. Only the reader was trained: under two hours on 8 H100s, on openly licensed English speech. The base model was never trained for speech-to-text at all.

  2. The speed receipt is specific, which is why I believe it. An hour of audio in 12.6 seconds on a base M3 MacBook Air with 8 GB, running on the Neural Engine: 288 times real time. On the same machine, IBM’s Granite Speech TurboCTC manages about 150x. The page also lists Parakeet numbers on larger newer Macs but flags them as a rough guide rather than a fair fight. Benchmarks with their own caveats attached earn trust.

  3. Using it is two gestures. Drop an audio file onto the app, or hold Right Option, speak, and release to dictate into whatever app is focused. Everything runs on the Neural Engine, so audio never leaves the machine. That is the entire pitch for anyone who dictates meeting notes they would rather not upload.

  4. The honest limitations are printed on the same page as the boasts. Word error rate sits around 5% against 2 to 3% for leading speech models, and the team says closing that gap is the next iteration. Vocabulary is English, roughly 229,000 words. Models download on first launch at 715 MB, so the 2.5 MB installer is just the shell. Mac only, Apple Silicon only, macOS 14 and up. If you need other languages or Windows, this is not your tool yet.

  5. This does what local Whisper and Moonshine setups do, through a cheaper and stranger door. Nobody trained a giant transcriber; somebody taught a small reader to eavesdrop on embedding space. And it fits Haku’s visible pipeline: the lab publishes small-model experiments, and the ones that work graduate into apps. Their $19 Dictate app already ships a 60M on-device formatting model from the same program.

Install & first run 

Open https://eg2.haku.sh/download in any browser.
You get EG2-Transcribe-1.6.dmg, about 2.5 MB.

That download resolves to Haku’s model host and fetched cleanly for me. Open it and drag the app to Applications like any Mac app.

Launch the app once: models download on first launch, 715 MB.
Drop in an audio file, or hold Right Option and speak into any app.

Success looks like a transcript appearing in seconds and dictation landing in your focused app on release. If the models have not finished downloading, transcription waits; that 715 MB is the real install.

Worth your time if: you transcribe or dictate on a Mac and want the fastest free offline option, accepting a few more errors than the best models. Skip it if you need non-English speech or anything beyond Apple Silicon.

Related TMFNK Content

Crepi il lupo! 🐺