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, an embedding model not trained for speech recognition, into a Mac speech-to-text tool. The project reports processing an hour of audio in 12.6 seconds on a base M3 MacBook Air, while its current word error rate is about 5%, compared with 2–3% for 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 tool for Apple Silicon Macs. The project says its app processed an hour of audio in 12.6 seconds on a base M3 MacBook Air, but reports a current word error rate of about 5%—higher than the 2–3% cited for leading speech models.
The method leaves EmbeddingGemma 2’s audio encoder unchanged. A separately trained reader maps short audio slices into the model’s text space, where the system selects the closest match from a vocabulary of about 229,000 English words. A small additional model supplies punctuation and capitalization. The project says the base embedding model itself was never trained to perform speech recognition.
According to the project report, training the reader took under two hours on eight H100 GPUs and used openly licensed English speech. The Mac app, listed as version 1.6, is free and requires macOS 14 or later on Apple Silicon. Its models download on first launch, with a stated download size of 715 MB. Users can transcribe an audio file or hold the Right Option key to dictate into an app.
The developer says transcription runs on the Mac’s Neural Engine, so audio stays on the device. In the project’s test on a base M3 MacBook Air with 8 GB of memory, it processed an hour of audio in 12.6 seconds, described as 288 times real time. The report also gives a 25.7-second result for IBM Granite Speech 5.0 Turbo on the same Mac, using the GPU through MLX. These are project-reported measurements, not an independent evaluation.
Fast Local Transcription, With an Accuracy Gap
The project shows how an existing multimodal embedding model can be repurposed for a task it was not built to handle, by training a smaller component to read information from its representation space. That approach may interest developers looking for ways to build speech tools from models that already encode audio and text in related forms. The report, however, does not establish that this technique will work as well across other languages, devices or recording conditions.
For Mac users, the main practical claims are high processing speed and local operation. Keeping audio on the computer may appeal to people who do not want recordings sent to a cloud service, while fast processing could make long recordings convenient to handle. Speed does not settle whether transcripts are reliable enough for a given use: the project’s own estimate of about 5% word error is above its cited 2–3% range for leading speech models. Accuracy may matter more than throughput for interviews, captions or records where incorrect words have consequences.
The reported comparison also needs care. The two M3 MacBook Air results were measured on the same machine, but the additional Parakeet and Phonon figures come from FluidAudio benchmarks on newer, larger Macs and different audio. The report calls those figures a rough guide, not a like-for-like ranking. Readers should not treat the figures as a controlled comparison across all the systems.
As an affiliate, we earn on qualifying purchases.
How the Embedding Model Reads Speech
EmbeddingGemma 2 is presented by its developer as a multimodal embedding model: inputs such as text and audio can be placed in a shared space, where related items land near one another. The project’s premise is that spoken words can therefore be associated with written words in that space, even though the original model was not trained to output a transcript.
Rather than retraining the audio encoder, the developer trained a reader to interpret its short audio representations as text. The application then chooses among English vocabulary entries and applies a separate punctuation-and-capitalization step. This differs from simply using an off-the-shelf speech-recognition model: the transcription behavior comes from the trained reader and decoding process, while the base model remains an embedding model.
The reported speed figures cover different hardware and test conditions. The 12.6-second result and the IBM comparison are from the project’s tests on an 8 GB M3 MacBook Air. Other listed results are attributed to FluidAudio’s published benchmarks on newer Macs and different audio. The source does not provide enough shared test details to compare every row directly.
“The base model was never trained for speech-to-text.”
— The project report
Accuracy and Test Conditions Still Matter
The report does not describe an independent test of the app’s 5% word error rate, including the test dataset, audio conditions or evaluation method. It also does not provide results broken down by accent, speaker, background noise or recording quality. The comparison with 2–3% for leading speech models is stated by the project, but the report does not name the evaluation basis for those figures.
The 12.6-second result is a project benchmark, and the source does not give enough detail to reproduce it or assess how performance changes with longer files, different formats or other Mac models. The broader benchmark table combines tests on different hardware and audio, so the listed speeds are not directly comparable across every system.
The report says training used openly licensed English speech and describes an English vocabulary. It does not establish support for other languages, explain the licensing status of every app component, or provide a privacy audit. The claim that audio stays on-device is the developer’s description of the app’s operation; the source does not offer independent verification.
A Better-Trained Reader Is Planned
The developer says the next iteration will focus on training the reader more effectively to reduce the word error rate. The report does not provide a release date, a target accuracy or a plan for independent testing, so it is not yet possible to say when or by how much the gap might narrow.
For now, readers can try the free version 1.6 app on an Apple Silicon Mac running macOS 14 or later, or wait for further accuracy and evaluation details. The next useful evidence would include documented test sets, results across varied speech and noise conditions, and comparisons run on the same hardware with the same audio.
Key Questions
What is the EmbeddingGemma 2 speech-to-text app?
It is a free Mac application that uses EmbeddingGemma 2 with a separately trained reader to turn audio into English text. The project says the original embedding model was not trained for speech-to-text.
How fast is the reported transcription?
The project reports processing one hour of audio in 12.6 seconds on a base M3 MacBook Air with 8 GB of memory, or 288 times real time. That is a developer-reported benchmark, not an independent result.
How accurate is the model?
The project reports a current word error rate of about 5% and compares it with 2–3% for leading speech models. It does not provide enough evaluation detail in the report to independently assess those figures.
Does audio leave the Mac?
The developer says transcription runs on the Mac’s Neural Engine and audio stays on the device. The source describes this as the app’s behavior but does not provide an independent privacy audit.
Which Macs can run the app?
The project lists support for Apple Silicon Macs running macOS 14 or later. Its first launch downloads models, with the project reporting a download size of 715 MB.
Source: hn
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
