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 transcribing an hour of audio in 12.6 seconds on a base M3 MacBook Air, but its word error rate is about 5%, compared with 2–3% for leading speech models.
A developer has adapted Google’s EmbeddingGemma 2, an embedding model not trained for speech-to-text, into a local transcription system for Macs. The project’s free app, listed as version 1.6, runs on Apple Silicon and reports processing one hour of audio in 12.6 seconds on a base M3 MacBook Air, while its current word error rate is about 5%.
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 closest match from roughly 229,000 English words. A second, small model adds punctuation and capitalization. The project says only the reader was trained, using openly licensed English speech in a run lasting under two hours on eight H100 GPUs.
On the project’s reported tests, the app processed an hour of audio in 12.6 seconds on an 8 GB base M3 MacBook Air, a stated speed of 288 times real time. The same machine processed the comparison system IBM Granite Speech 5.0 Turbo in 25.7 seconds, or about 150 times real time. Those figures are the project’s tests, not an independently verified evaluation.
The app accepts an audio file or supports dictation through a keyboard shortcut: hold Right Option while speaking, then release. The developer says processing runs on the Mac’s Neural Engine and audio does not leave the device. Models, totaling 715 MB according to the project page, download on first launch. The app requires macOS 14 or later and Apple Silicon.
Fast Transcription Without Uploads
The demonstration points to a different use for embedding models: their shared representations of audio and text can support transcription when paired with a reader trained to recover words. The base model itself was not trained to transcribe speech, so the project is evidence of a particular adaptation rather than a general capability established for all embedding models.
Local processing may appeal to people who want to dictate or transcribe recordings without sending audio to a remote service. It also makes performance dependent on compatible hardware and the downloaded models. The reported speed is striking, but speed alone does not establish transcription quality or reliability across accents, recording conditions, languages, and specialist vocabulary.
For practical use, the key trade-off is the project’s stated roughly 5% word error rate. The report compares that with 2–3% for leading speech models, meaning the prototype’s accuracy currently trails those systems even as it runs quickly on a laptop. The comparison supports cautious interest, not a claim that the app is already a replacement for established transcription tools.
As an affiliate, we earn on qualifying purchases.
How The Reader Uses Embeddings
Embedding models represent different inputs as vectors so that related items can sit near one another in a shared mathematical space. The project says EmbeddingGemma 2 supports text, images, audio and video; spoken words can therefore be represented near their written counterparts. Its approach reads from that space rather than replacing or retraining the base audio encoder.
The project’s benchmarks require careful comparison. Its M3 MacBook Air results for the new app and IBM Granite Speech were measured on the same machine. Other figures on the project page come from FluidAudio’s published tests of Parakeet-based models on larger, newer Macs and different audio. The developer explicitly describes those other rows as a rough guide, not a like-for-like ranking.
The reported training was limited in scope: under two hours on eight H100 GPUs using openly licensed English speech. The page does not describe the training set’s size or composition in detail. The app is presented as an English-word transcription system, and the supplied material does not establish support for other languages.
“The base model was never trained for speech-to-text.”
— The EmbeddingGemma 2 transcription project
on-device speech transcription Mac
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Accuracy And Test Conditions
The project gives a current word error rate of about 5%, but the supplied report does not specify the test set, its size, the audio conditions, or the evaluation method. Without those details, readers cannot determine how the result might vary by accent, speaking style, background noise, or recording quality.
The speed figures are project-reported. No independent replication is provided, and comparisons involving FluidAudio models use different hardware and audio, so they do not establish a controlled performance ranking. The privacy statement also comes from the developer; the source material does not include an external technical review of network behavior.
It remains unclear how the reader performs outside the described English use case, how well it handles names and uncommon words, and whether a new training run will reduce errors. The developer says a better-trained reader is the next iteration, but provides no release date or projected accuracy.
As an affiliate, we earn on qualifying purchases.
A Better-Trained Reader
The project identifies training a better reader as its next step and says it believes that work could narrow the accuracy gap with leading speech models. That is a development goal, not a demonstrated result; no timetable, target word error rate, or details of the planned training data are given.
For now, the app is available as version 1.6 for Apple Silicon Macs running macOS 14 or later, with models downloaded on first launch. Future updates and more detailed evaluations would help establish whether the current speed and on-device design can be paired with improved accuracy across a broader range of recordings.
As an affiliate, we earn on qualifying purchases.
Key Questions
What does the EmbeddingGemma 2 app do?
It turns audio into English text on a compatible Mac. The developer says users can choose an audio file or dictate into an app using a keyboard shortcut.
Was EmbeddingGemma 2 originally trained to transcribe speech?
No. The project says the base model was not trained for speech-to-text. It keeps the audio encoder unchanged and trains a small reader to map audio slices into the model’s text space.
How accurate is the transcription?
The developer reports a word error rate of about 5% and compares it with 2–3% for leading speech models. The source does not provide enough information about the test set or evaluation method to independently assess that comparison.
Does the audio get uploaded?
The project says the app runs on the Mac’s Neural Engine and that audio stays on the device. That is the developer’s description; the supplied material does not include an independent privacy audit.
Which Macs can run it?
The app is listed for Apple Silicon Macs running macOS 14 or later. The project says its models, totaling 715 MB, download when the app is first launched.
Source: hn
Columbus Day / Indigenous Peoples' Day Picks
long weekend sales
As an affiliate, we earn on qualifying purchases.
