TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
A project called EmbeddingGemma Transcribe uses a small trained reader to turn audio representations from Google’s EmbeddingGemma 2 into text. Its creators report that it processes an hour of audio in 12.6 seconds on a base M3 MacBook Air, but say its current word error rate is about 5%, above the 2–3% they cite for leading speech models.
The creators of EmbeddingGemma Transcribe say they have adapted Google’s EmbeddingGemma 2, a model built to create shared representations of text, images, audio and video, into a speech-to-text system that runs locally on Apple Silicon Macs. Their project page reports that a base M3 MacBook Air can process one hour of audio in 12.6 seconds; the reported word error rate is about 5%, higher than the 2–3% the creators give for leading speech models.
The system leaves EmbeddingGemma 2’s audio encoder unchanged. A separately trained, small reader maps short audio slices into the model’s text space. The system then selects the nearest match from roughly 229,000 English words, combining the slices into a transcript. A small additional model supplies punctuation and capitalization.
The project report says only the reader was trained, in under two hours on eight H100 GPUs, using openly licensed English speech. The creators stress that the underlying EmbeddingGemma model was not trained for speech recognition. The project offers a free Mac download listed as version 1.6; the models download on first launch, and the reported download size is 715 MB. It requires macOS 14 or later and Apple Silicon.
For its speed comparison, the team says its first two results were measured on the same 8 GB M3 MacBook Air. EmbeddingGemma Transcribe processed an hour in 12.6 seconds, while IBM Granite Speech 5.0 Turbo took 25.7 seconds in the team’s test. The project also lists faster results for other systems on newer Macs, but those figures come from FluidAudio’s published benchmarks and use different machines and audio. The report cautions that the results are not a like-for-like ranking.
Fast Transcription Without Cloud Uploads
The project’s appeal is the combination of local processing and speed. Its creators say the app runs on the Mac’s Neural Engine and audio does not leave the computer. That could suit people who want quick dictation or transcription without sending recordings to an online service, although the report does not provide an independent privacy audit or describe every data-handling detail.
Its reported accuracy is a meaningful limitation. A word error rate of about 5% means the system is not presented as matching the 2–3% rate the team attributes to leading speech models. Errors can matter in interviews, records or other work where names and exact wording need checking. The speed result is promising, but it does not by itself establish suitability for every transcription task.
The experiment also shows how an embedding model might be repurposed: rather than training the base model to generate transcripts, the developers use its existing audio and text representations and train a compact component to connect them. That is the project’s method, not evidence that embedding models generally can replace speech-recognition systems.
Apple Silicon compatible speech-to-text software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Embedding models typically represent different inputs as vectors so that related items sit near one another in a mathematical space. The project’s premise is that if spoken audio and its written words are close in EmbeddingGemma 2’s shared multimodal space, a reader may be able to recover likely words from the audio representations. The transcription method uses that nearest-word approach, rather than asking the base model to produce text directly.
The speed figures need to be read alongside their test conditions. The team’s comparison between EmbeddingGemma Transcribe and IBM Granite Speech used the same base M3 MacBook Air, according to the report. Its other listed comparisons include Parakeet-based systems tested on M4 Pro and M5 Pro machines using FluidAudio’s published results. Different hardware and audio make those numbers rough guides, not controlled comparisons.
“The base model was never trained for speech-to-text.”
— EmbeddingGemma Transcribe project report
local speech recognition app for MacBook Air
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Accuracy and Testing Still Need Review
The reported 5% word error rate is not accompanied in the supplied project material by details such as the test-set size, accents represented, recording conditions or evaluation method. It is therefore not possible from that report alone to judge how performance changes across speakers, noisy recordings, languages or specialist vocabulary. The project describes the system as using English words and openly licensed English training speech; broader language support is not established.
The speed figures are also project-reported rather than an independent, comprehensive evaluation. Only the M3 MacBook Air comparison is described as testing both systems on the same computer, while other listed results use different hardware and sources. The report says a better-trained reader may reduce the accuracy gap, but gives no measured result or schedule for that work. How the app performs on long, varied recordings and how consistently it handles punctuation remain unclear.
offline transcription software for Mac
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The Reader Is the Next Target
The creators say the next iteration will focus on training a better reader to improve accuracy. They have not supplied a release date or a target error rate. Users can currently try the listed version 1.6 on compatible Macs, with models downloaded at first launch; the project describes both file transcription and a hold-to-dictate shortcut using the Right Option key.
Further evaluation would help establish whether the current speed and accuracy hold across a broader set of recordings and whether later reader training narrows the reported gap with leading speech models. Until such results are published, the project’s numbers should be treated as the developers’ reported benchmarks, not a general guarantee of performance.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is EmbeddingGemma Transcribe?
It is a project that adapts Google’s EmbeddingGemma 2 for speech-to-text by adding a trained reader that maps audio slices to nearby words in the model’s text space.
How fast does it transcribe audio?
The project reports that a base M3 MacBook Air with 8 GB of memory processes one hour of audio in 12.6 seconds. This is a reported benchmark, not a guarantee for every recording or Mac.
How accurate is the current system?
The creators report a word error rate of about 5%. They cite 2–3% for leading speech models, but the supplied report does not give enough test-set detail to independently compare those figures.
Does audio go to an online service?
The project says transcription runs on the Mac’s Neural Engine and that audio stays on the device. The report does not include an independent audit of the app’s data handling.
Which Macs can run it?
The download is listed for Apple Silicon Macs running macOS 14 or later. The project says its models download on first launch and take 715 MB.
Source: hn
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
