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 developer has adapted Google’s EmbeddingGemma 2 embedding model for speech-to-text in a free Mac app that processes audio on-device. The project reports roughly 5% word error and fast processing on a base M3 MacBook Air, but its comparisons are not all like-for-like and accuracy remains behind leading speech models.
A developer has adapted Google’s EmbeddingGemma 2, a model designed to map text, images, audio and video into a shared embedding space, into a speech-to-text app for Apple Silicon Macs. The project says its app transcribes an hour of audio in about 13 seconds on a base M3 MacBook Air and keeps processing on the device, while reporting a word error rate of about 5%.
The app, described on the project website as Embedding Gemma Transcribe, is offered free for macOS 14 and later. Its 715 MB of model files download the first time the app launches. Users can provide an audio file or dictate into another app by holding the Right Option key while speaking and releasing it when finished, according to the project.
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 selects the closest match from a vocabulary of about 229,000 English words and combines the slices into a transcript. A second, small model adds punctuation and capitalization.
The project says only the reader was trained, in less than two hours on eight H100 GPUs using openly licensed English speech. Its reported speed on an 8 GB base M3 MacBook Air is 288 times real time: one hour of audio took 12.6 seconds. The report says the app runs on the Mac’s Neural Engine and that audio does not leave the computer.
Fast Local Transcription, With an Accuracy Trade-Off
The project offers an example of using a model for a task beyond its original design: an embedding model that represents relationships between different kinds of data is being used as the basis for word recognition. That may interest developers exploring whether existing multimodal models can support practical applications without replacing their core encoders.
For users, the reported combination of on-device processing and high speed could suit private dictation or transcription of long recordings. Local processing means the project says audio is not sent to a cloud service, although users still need to assess the app and their own device and data practices.
The main qualification is accuracy. The report puts the system’s word error rate at about 5%, compared with 2% to 3% for leading speech models. A faster transcript is not necessarily a more useful one, particularly for names, technical vocabulary or other material where errors matter. The developer says a better-trained reader may narrow the gap, but presents that as a goal for a future iteration, not a result already achieved.
As an affiliate, we earn on qualifying purchases.
How an Embedding Model Reads Audio
EmbeddingGemma 2 is presented by the project as a multimodal embedding model: audio and written words can occupy nearby positions in a shared representational space. The developer’s approach uses that relationship to identify likely words from audio rather than training the base model as a conventional speech-to-text system.
The reported speed comparisons need careful reading. The project tested Embedding Gemma Transcribe and IBM Granite Speech 5.0 Turbo on the same base M3 MacBook Air; it reports 12.6 seconds and 25.7 seconds, respectively, to process one hour of audio. Other listed figures come from FluidAudio’s published tests of Parakeet-based models on larger, newer Macs and different audio. The project explicitly describes those results as a rough guide, not a like-for-like ranking.
The app’s first-launch download is listed as 715 MB, and use is limited to English in the described method, whose vocabulary is an English word list. The report does not establish performance across accents, recording conditions, or specialist terms.
“The base model was never trained for speech-to-text.”
— The project report
on-device transcription software for Mac
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Accuracy Across Real-World Audio
The published material does not provide details about the test set, its size, or the calculation behind the roughly 5% word error rate. It also does not establish how results change with different accents, background noise, overlapping speakers, or specialized vocabulary. The accuracy figure should therefore be treated as the project’s report, not an independently verified benchmark.
Some speed results are direct comparisons on the same Mac, while others use different devices, audio and testing sources. Those numbers cannot establish a universal performance ranking. It is also not clear from the available report how the app behaves on Mac configurations beyond the cited test machine or whether later releases will change its model, performance or accuracy.
privacy-focused speech recognition Mac
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
A Better-Trained Reader Is Planned
The developer identifies training a better reader as the next iteration, with the stated aim of reducing the gap between the app’s reported word error rate and leading speech models. The report does not give a release date, target accuracy, or details about planned training data, so the timing and likely improvement remain unknown.
For now, the app is available as version 1.6 for supported Macs, with model files downloaded on first launch. Further public evaluations that describe their datasets, test conditions and accuracy calculations would help users compare it with established transcription systems. Until then, its reported speed and local processing are project claims, while its accuracy limitations warrant caution for transcripts that require high precision.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Embedding Gemma Transcribe?
It is a Mac speech-to-text app built around Google’s EmbeddingGemma 2. The project says a small trained reader turns audio representations into words and another small model adds punctuation and capitalization.
Does the app send audio to a server?
The project says transcription runs on the Mac’s Neural Engine and audio never leaves the device. That describes the developer’s stated design; the report does not provide an independent privacy audit.
How accurate is its transcription?
The project reports a word error rate of about 5%, compared with 2% to 3% for leading speech models. It does not publish enough testing detail in the supplied report to independently assess that comparison.
How fast does it process audio?
On an 8 GB base M3 MacBook Air, the project reports processing one hour of audio in 12.6 seconds. Its other comparisons use different machines and test conditions, so they are not directly equivalent.
Which Macs and languages are supported?
The project lists macOS 14 or later and Apple Silicon. Its described system uses a vocabulary of about 229,000 English words, and the report does not document support for other languages.
Source: hn
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
