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 Embedding Gemma Transcribe adapts Google’s multimodal embedding model for on-device speech transcription on Apple Silicon Macs. Its creators report processing an hour of audio in 12.6 seconds on a base M3 MacBook Air, while saying its current word error rate is about 5%, above the 2–3% they cite for leading speech models.
A project has adapted Google’s EmbeddingGemma 2, a model designed to map text, images, audio and video into a shared representation space, into a speech-to-text tool for Apple Silicon Macs. Its creators say the app transcribes locally, with a test processing one hour of audio in 12.6 seconds on a base M3 MacBook Air, though they report a current word error rate of about 5%.
The project, presented as Embedding Gemma Transcribe, does not retrain EmbeddingGemma 2 itself as a conventional speech-recognition model. According to its report, the model’s audio encoder remains unchanged. A separately trained, small reader maps short audio slices into the model’s text space; the system then selects the closest match from a vocabulary of roughly 229,000 English words. A small additional model supplies punctuation and capitalization.
The creators say only the reader was trained, in under two hours on eight H100 GPUs, using openly licensed English speech. They say the underlying embedding model was not trained for speech-to-text. The downloadable Mac app is listed as version 1.6, free to use, compatible with macOS 14 or later and Apple Silicon, and approximately 715 MB; model files download when the app is first opened.
For speed, the project reports a result of 12.6 seconds per hour of audio on a base M3 MacBook Air with 8 GB of memory, using the Mac’s Neural Engine. Its comparison table also lists IBM Granite Speech 5.0 Turbo at 25.7 seconds per hour on the same Mac, with GPU processing through MLX. The project describes these as its own tests. Other table entries come from FluidAudio’s published benchmarks on different, newer Macs and audio, so the report warns that they are not a like-for-like ranking.
Fast Local Transcription, With an Accuracy Gap
The project is notable because it uses a model built for multimodal embeddings as the basis for transcription, rather than replacing its audio encoder with a conventional speech-recognition system. That approach may be useful to developers exploring how shared audio-and-text representations can support tasks beyond the model’s original role. The report, however, does not establish that this method matches dedicated transcription models on accuracy.
For Mac users, the proposed benefit is a combination of speed and local processing. The creators say audio stays on the computer, which avoids sending recordings to a remote service during transcription. That could matter for private recordings or users working without a network connection, although the report does not describe an independent privacy or security assessment.
Accuracy remains the main trade-off. The creators put the system’s current word error rate at about 5%, compared with 2–3% for leading speech models. Those figures are claims in the project report; it does not provide enough evaluation detail here to establish how the systems were tested or whether they used the same audio and scoring conditions. A faster result is not by itself evidence of more accurate transcription.
As an affiliate, we earn on qualifying purchases.
How the Embedding Approach Works
EmbeddingGemma 2 represents different kinds of content in a shared space, where related inputs can be placed near one another. The project’s stated premise is that spoken words and their written forms may land close enough in that space for a reader to identify a likely word from audio features. Rather than generating text through a standard speech-to-text decoder, this system chooses among candidate words and then adds basic written formatting.
The published workflow is designed for direct use on a Mac: users can provide an audio file or hold the Right Option key, speak, and release the key to dictate into an app. The project says processing runs on the device’s Neural Engine. Its performance figures should be read within the stated hardware and test limits: the direct comparison is between two systems tested on a base M3 MacBook Air, while the remaining table results involve other machines and benchmarks.
““The base model was never trained for speech-to-text.””
— Embedding Gemma Transcribe project report
local speech recognition app for Mac
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Accuracy Tests and Privacy Details
The report gives an approximate 5% word error rate, but does not specify the test set, its size, recording conditions, language and accent coverage, or the exact scoring procedure. It also does not provide enough information to verify that the cited 2–3% rates for leading models were measured on the same data. The accuracy comparison should therefore be treated as the project’s own summary, not a controlled independent benchmark.
It is also unclear how well the word-selection method handles names, uncommon vocabulary, accents, overlapping speakers, and noisy recordings. The report describes English speech and a fixed vocabulary, but does not give detailed results across these cases. Its local-processing statement explains where the project says transcription happens; it does not provide a full account of what data the app stores or how every component behaves.
Apple Silicon speech transcription tool
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
A Better-Trained Reader Is Planned
The project says its next iteration will focus on training a better reader in an effort to reduce the gap between its current reported word error rate and those of leading speech models. It has not supplied a release date, a target accuracy figure, or a detailed evaluation plan for that work. For now, version 1.6 is available for compatible Apple Silicon Macs, and the next meaningful measure will be whether later results show improved accuracy on clearly described tests without sacrificing local processing or speed.
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. Its creators say a separately trained reader maps audio into the model’s text space to select words.
Does it process audio in the cloud?
The project says transcription runs on the Mac’s Neural Engine and that audio does not leave the machine. That statement comes from the project report; no independent privacy audit is included in the supplied material.
How fast is the reported transcription?
The creators report processing one hour of audio in 12.6 seconds on a base M3 MacBook Air with 8 GB of memory. They also warn that results from other Macs in their comparison table use different hardware and tests.
How accurate is it compared with leading speech models?
The project reports a current word error rate of about 5% and cites 2–3% for leading speech models. The report does not provide enough test details to confirm that those figures come from comparable evaluations.
Which Macs can run the app?
The project lists macOS 14 or later and Apple Silicon as requirements. Its report lists version 1.6 as a free download and says model files download on first launch.
Source: hn
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
