Transcribe an Hour of Audio in 13 Seconds on Your Mac—No Cloud, No Costs
A research project released this week demonstrates that a multimodal embedding model can double as a speech-to-text engine—no retraining of the base model required. The system, called Embedding Gemma Transcribe, processes an hour of audio in roughly 13 seconds on a base M3 MacBook Air with 8 GB of memory. It runs entirely on Apple's Neural Engine. Audio never leaves the device.
How this even works
Google's EmbeddingGemma 2 maps text, images, audio, and video into a single vector space. A spoken word and its written form land near each other. The researchers exploited this property: they kept the audio encoder frozen and trained only a small "reader" model that projects short audio slices into the text embedding space. The closest match among roughly 229,000 English words becomes the transcript token. A separate tiny model adds punctuation and capitalization.
The elegance matters: training the reader took under two hours on eight H100 GPUs using openly licensed English speech data. The base EmbeddingGemma 2 model was never trained for speech recognition. This is a creative repurposing, not a from-scratch build.
What creators are actually doing with it
The response on Hacker News and the project site shows real use cases. One tester noted they processed 33 minutes of local audio on an M4 Mac Mini in 6.75 seconds—competitive with existing tools but without leaving the machine. For podcasters managing long interviews, video editors transcribing raw footage, and researchers dealing with audio archives, the math is simple: this replaces a subscription cost with a one-time download.
The free Mac app lets you drop audio files directly or hold Right Option to dictate into any application. Researchers are already extending it; the use of eight H100 GPUs for two hours transformed Google's EmbeddingGemma 2 into a state-of-the-art tool for Vietnamese speech recognition after training on 3,000 hours of audio. This development expands the utility of the embedding model into speech-to-text tasks, following another community project that repurposed the model to transcribe one hour of audio in about 13 seconds on a base M3 MacBook Air.
The honest limits
Benchmarks on the same base M3 MacBook Air show Embedding Gemma Transcribe at 288× real time (12.6 seconds per hour of audio). Word error rate sits around 5 percent. Leading dedicated speech models achieve 2–3 percent. The team believes a better-trained reader can close the gap and plans to iterate. That 5% error rate is the trade-off: it's good enough for drafts and for creators who don't need perfection, but not for legal or medical transcription.
One tester on Hacker News noted that it transcribes Portuguese audio into English (not native Portuguese support yet), which is useful for some workflows but a limitation to keep in mind. Punctuation and capitalization are added by a separate model, so expect some rough edges there too.
How to try it
The Mac app is free and requires macOS 14 or later on Apple Silicon. Models (715 MB) download on first launch. You can download it at eg2.haku.sh. If you want to experiment with the underlying models on other systems, this reader is an independent reproduction, trained from scratch on LibriSpeech. Weights are available on Hugging Face.
For creators working entirely offline or managing transcription costs at scale, this shifts the calculus. You're no longer paying per minute; you're running it once on your hardware. That matters if you're processing hundreds of hours a year.