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 report describes adapting Google’s EmbeddingGemma 2, an embedding model rather than a speech recognizer, to transcribe audio on Apple Silicon Macs. The developer reports processing an hour of audio in 12.6 seconds on a base M3 MacBook Air, while giving the system a current word error rate of about 5%, compared with 2–3% for leading speech models.
A developer has adapted Google’s EmbeddingGemma 2, an embedding model not designed for speech recognition, into an on-device transcription tool for Apple Silicon Macs. The project report says it processed one hour of audio in 12.6 seconds on a base M3 MacBook Air, but its reported word error rate of about 5% remains higher than the 2–3% cited for leading speech models.
The project, presented at eg2.haku.sh as Embedding Gemma Transcribe, keeps EmbeddingGemma 2’s audio encoder unchanged. Its developers trained a small additional component, called a reader, to move short audio slices into the model’s text space. The system selects the closest match from a vocabulary of about 229,000 English words; a separate, small model adds punctuation and capitalization.
The report says only the reader was trained, in under two hours on eight H100 GPUs, using openly licensed English speech. It emphasizes that the base model itself was not trained for speech-to-text. The Mac app is listed as free, version 1.6, and requires macOS 14 or later on Apple Silicon. It downloads model files on first launch, with a stated download size of 715 MB.
For its performance comparison, the project reports that its system ran at 288 times real time on an 8 GB base M3 MacBook Air, transcribing an hour in 12.6 seconds. On the same computer, IBM Granite Speech 5.0 Turbo took 25.7 seconds, or about 150 times real time. The source cautions that other listed results come from FluidAudio benchmarks on newer, more powerful Macs and different audio, so they are not a like-for-like ranking.
Fast Transcription Without Cloud Uploads
The project’s reported speed and local processing could appeal to Mac users who want to transcribe recordings without sending audio to a cloud service. The developer says processing runs on the Mac’s Neural Engine and that audio never leaves the machine. That is a project description of how the app operates; the source does not provide an independent privacy or security audit.
The approach also demonstrates a different use for an embedding model. Rather than generating text through a conventional speech-recognition system, it uses relationships between audio and text representations to select words. The reported result is promising on speed, but the stated 5% word error rate indicates a meaningful accuracy gap against the leading systems cited by the developer. For users, the trade-off is therefore not simply speed: accuracy on their own recordings and the value of local processing will matter too.
Mac speech-to-text transcription software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
EmbeddingGemma 2 is described by the project as a multimodal embedding model that places text, images, audio and video in a shared representational space. The idea behind the adaptation is that spoken audio and its corresponding written word can land near each other in that space. The trained reader is intended to turn that relationship into a sequence of likely words.
This differs from retraining the underlying model as a speech recognizer: according to the report, the encoder stays as it was and only the reader is trained. The project’s comparison table includes results from other speech models, but combines its own tests on an M3 MacBook Air with published FluidAudio figures from M4 Pro and M5 Pro machines. The source explicitly describes the figures as a rough guide, not a controlled comparison across identical hardware and audio.
As an affiliate, we earn on qualifying purchases.
Accuracy and Testing Still Need Clarity
The report does not specify the test dataset, recording conditions, or evaluation procedure behind its approximately 5% word error rate. It also does not give the test details behind the cited 2–3% range for leading speech models. The figures therefore offer a useful stated comparison, but readers cannot determine from the source how closely the tests match their own languages, accents, background noise or recording quality.
The speed figures have limits too: only the M3 MacBook Air results are described as tests on the same computer, while other entries use different Macs and audio. The report does not provide independent validation of the performance, privacy claims or transcription quality. It also describes English-word matching; support for other languages is not established. Whether a better-trained reader will reduce errors remains an expectation from the project, not a confirmed improvement.
privacy-focused speech transcription Mac
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
A More Accurate Reader Is Planned
The project says its next iteration will focus on training a better reader to reduce the reported error rate. It does not provide a release date, a target accuracy, or details about how that training will differ. For now, readers can download the listed Mac app, subject to its stated macOS 14 and Apple Silicon requirements, and assess its results on their own material. Further independent testing and clearer benchmark methods would help establish how its speed and accuracy compare under consistent conditions.
As an affiliate, we earn on qualifying purchases.
Key Questions
Is EmbeddingGemma 2 itself a speech-to-text model?
No. The project describes EmbeddingGemma 2 as an embedding model that was not trained for speech-to-text. A separately trained reader maps its audio representations to likely words.How fast does the project say it transcribes audio?
The report says it processed one hour of audio in 12.6 seconds on an 8 GB base M3 MacBook Air, or 288 times real time. This is the developer’s reported test result.How accurate is the transcription?
The project reports a word error rate of about 5%, compared with 2–3% for leading speech models. The source does not provide enough testing detail to independently assess that comparison.Does the audio go to a cloud service?
The project says the app runs on the Mac’s Neural Engine and that audio stays on the device. The supplied report does not include an independent privacy audit.Which Macs can run the app?
The listing specifies Apple Silicon and macOS 14 or later. It says model files download on first launch and total 715 MB.Source: hn
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
