TL;DR
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
A developer says its experimental Mac transcription app processed one hour of audio in 12.6 seconds on a base M3 MacBook Air, running locally on the Neural Engine. The reported word error rate is about 5%, compared with 2–3% for leading speech models, and the developer says accuracy work is still needed.
A developer says its free EmbeddingGemma Transcribe app processed an hour of audio in 12.6 seconds on a base M3 MacBook Air with 8 GB of memory, running on the Mac’s Neural Engine. The reported speed comes with a current word error rate of about 5%, higher than the 2–3% cited for leading speech models, so the result is a performance claim rather than evidence of comparable transcription accuracy.
The app is built around Google’s EmbeddingGemma 2, a multimodal embedding model that maps text, images, audio and video into a shared space. According to the developer’s explanation, the underlying model was not trained as a speech-to-text system. Instead, a separately trained reader maps short audio slices into the model’s text space and selects the closest match from a vocabulary of about 229,000 English words. A small additional model supplies punctuation and capitalization.
The developer says only that reader component was trained, in under two hours on eight H100 GPUs, using openly licensed English speech. The original audio encoder was left unchanged. The app’s download is listed as 715 MB, with models downloaded on first launch; it requires macOS 14 or later and Apple Silicon. The developer says processing happens locally and audio does not leave the Mac.
The report compares the app’s result with IBM Granite Speech 5.0 Turbo, which it says processed an hour in 25.7 seconds on the same base M3 MacBook Air. That comparison is presented as a test by the app’s developers. Other figures on the page come from FluidAudio benchmarks on newer, larger Macs and different audio, and the source cautions that these are not like-for-like results.
Speed Comes With an Accuracy Trade-Off
If the reported timing holds across ordinary recordings, local transcription at this speed could make long files practical to process on a laptop without sending audio to a cloud service. The app’s use of the Neural Engine also illustrates a different approach from building a conventional speech-recognition model: it repurposes an embedding model’s relationship between sound and text, with a small reader translating audio representations into words.
For users, however, speed is only one part of the result. A reported 5% word error rate means the system may still misrecognize words, which can matter for names, technical vocabulary and records that need high accuracy. The source contrasts that figure with 2–3% for leading speech models, but does not name those systems or describe a shared test set. The published numbers therefore do not establish an independent, directly comparable accuracy ranking.
On-device processing may also appeal to people handling private recordings, because the developer says audio stays on the computer. That is a statement about this app’s stated operation, not a third-party security audit. Users with sensitive material would still need to evaluate the app and its permissions for themselves.
MacBook Air audio transcription software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
How the Embedding Model Reads Speech
Embedding models are designed to represent different kinds of content in a shared mathematical space, where related items can sit near one another. The developer’s proposal is that spoken audio and its written form can be close enough in EmbeddingGemma 2’s representation for a reader to infer words from audio slices, even though the base model was not trained for transcription.
The reported system is thus not simply a conventional speech-recognition model running faster. It combines the unchanged audio encoder with a learned reader and a separate punctuation-and-capitalization component. The developer describes the reader’s training data as openly licensed English speech, which identifies the language and broad data source but does not provide a detailed account of the training corpus or evaluation methodology.
The source lists a one-hour processing time of 12.6 seconds for the app on an 8 GB M3 MacBook Air, equivalent to roughly 288 times real time. Its table also includes tests or published benchmarks for other models and hardware, but explicitly warns that results from newer Macs and different audio are only a rough guide. Those entries should not be read as a controlled comparison across all products.
““The base model was never trained for speech-to-text.””
— The app’s developer
on-device speech-to-text app for Mac
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Accuracy Testing Still Needs Detail
The source does not specify the date of the speed test, the recording’s length beyond the one-hour benchmark, its audio conditions, or whether the timing includes file loading and setup. It also does not provide a test set, sample size or methodology for the about 5% word error rate, nor identify the leading models said to score 2–3%. Those omissions make it difficult to compare accuracy or determine how performance changes with accents, background noise, speaker count and specialist vocabulary.
The reported result comes from the developer’s own page, and no independent replication is included in the supplied material. The claim that recordings remain on-device is also attributed to the developer; the source does not describe an external audit. It is not yet clear how the app performs on languages other than English, or whether the speed and accuracy figures generalize beyond the tested setup.
Mac neural engine transcription tool
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
A More Accurate Reader Is Planned
The developer says the next iteration will focus on improving the reader, arguing that better training could reduce the current accuracy gap. No release date, target word error rate or testing plan is given. The app is currently listed as a free download, version 1.6, for Apple Silicon Macs running macOS 14 or later.
The next useful evidence would include independently repeatable timing tests and accuracy results on a clearly described dataset, with comparisons made under the same conditions. Until those details are available, the 12.6-second figure is a notable developer-reported speed result, while the app’s practical accuracy remains a key open question.
As an affiliate, we earn on qualifying purchases.
Key Questions
How long did the app take to transcribe an hour of audio?
The developer reports 12.6 seconds on a base M3 MacBook Air with 8 GB of memory. The supplied material does not specify the audio conditions or whether setup time was included.
How accurate is EmbeddingGemma Transcribe?
The developer reports a word error rate of about 5%. It cites 2–3% for leading speech models, but does not name them or provide a shared evaluation method, so the comparison is not independently verifiable from the source.
Does the transcription run in the cloud?
The developer says the app runs on the Mac’s Neural Engine and that audio does not leave the computer. The source does not include an independent privacy audit.
What Macs can run the app?
The app is listed for Apple Silicon Macs running macOS 14 or later. Its models download on first launch, and the listed download is 715 MB.
Is the underlying model a speech-recognition model?
No. The developer says EmbeddingGemma 2 was not trained for speech-to-text. The app adds a trained reader that maps short audio slices into the model’s text space, plus a small component for punctuation and capitalization.
Source: hn
Columbus Day / Indigenous Peoples' Day Picks
long weekend sales
As an affiliate, we earn on qualifying purchases.
