Google just released a Mac version of its on-device AI dictation app Eloquent. The app allows you to choose between two Gemma 4 local models: Gemma 4 2B (the 2 billion parameter model) and Gemma 12B (12 billion parameters). The latter is designed to run on Apple silicon Macs with at least 16gb of ram, i.e., just about anything but the Neo.
It bills itself as a free dictation app with premium features, e.g., polishing the text to remove disfluencies and improve flow, user-added custom vocabulary, a global hotkey so you can use it in an open app, etc. (You can toggle off the polishing if you wish.) You can also upload an audio file for it to transcribe. You can ask it to turn your text into bullet points or a summarizy of key points or revise the text to adjust your tone. You can highlight text youâve already written and, via a voice command, ask it to restructure or adjust the tone of the text.
Iâm test-driving it now and it is more than adequate for my needs. Itâs fast and its on-the-fly polishing to eliminate filler words and improve flow works well. I havenât tried any of the other features yet.
A whole category of subscriptions rendered obsolete (hopefully). Huzzah!
FWIW I already use a one time payment app but Iâve seen many people donât.
Available first on iOS and macOS. lol.
I think the name is Google AI Edge Eloquent, which is quite a mouthful. Pun not intended but gladly acceptedâŚ
Iâd be interested to see if it has dictionary/saved corrections functionality.
Thatâs the name on its birth certificate. Fortunately, it installs itself as Eloquent. In the grand tradition of Garbo, Bogart, Mozart, and Dylan âŚ
You can create your own dictionaries. (Note the plural: you can created collections grouped by categories of your choosing.) There is a pre-defined category named Learned from edits."
I assume that there will be a market for installable, pre-defined specialist dictionaries if there isnât one already.
So, I just tested Eloquentâs ability to transcribe an mp3 audio track extracted from a 27 minute YouTube video. (This one.)
It appeared to have no problem creating the transcription itself, although I could see from the raw transcript that there were some glitches. Those were corrected in the polishing phase, during which it also divided the text into paragraphs and added some punctuation. (Yes! EM dashes!) But ⌠It also made some changes to smooth out the kind of infelicities that creep into a spoken text, even when itâs as carefully scripted as this YouTube video was, e.g. the random âyou know.â Some of the changes were of the kind an editor might make to a text headed for print. For example: âThese columns are what determines what vector each word turns into in that first stepâ was changed to âThese columns determine what vector each word turns into in that first step.â
It would appear that Eloquentâs goal is to polish something spoken into something ready for print. So long as you understand that going in, itâs a nice tool to have in the toolbox. You do always have the option to use the raw transcript as is if you prefer.
Note: Iâm using an M2 Mac Studio with 64gb of ram.
I feel happy for us all (assuming the dream turns into reality * ).
But I feel sorry for the developers whoâve invested so much time and money into the really cool transcription apps.
I know thatâs how capitalism works, but it must be a miserable being an indie developer trying to make a living.
fn * I doubt Google will put the heavy lifting into making their app nice and easy and clean to use. Theyâre good at big heavy stuff, but not the subtle, tasteful kind.
In terms of ease-of-use, itâs actually ⌠not bad.
It might be all a casual user like me needs, but I donât think itâs going to put the big dictation apps like Whispr Flow out of business anytime soon. For one thing, the only language it currently supports is English. Itâs almost certainly not HIPAA compliant, and Iâm going to hazard a guess that it doesnât do the kind of code syntax highlighting that Whispr Flow does.
And note that while it handles basic dictation just fine, itâs not a fully-featured voice-to-text transcription app. For instance, it canât provide speaker diarisation on uploaded audio files. For that, youâll need another solution. The paid version of the Gemini app itself does this beautifully and can handle very large audio files, but there are also plenty of purpose-built alternatives.
If you use the Mac version of the Eloquent app, everything happens locallyâand is stored locallyâon your Mac using a locally installed Gemma 4 model. (You have to download a copy of the model before you can use the app.) Nothing gets sent to Googleâs servers.
The iOS version of the app, which uses a much smaller Gemma model, does have a toggle that will put the app in Cloud mode to enable the premium features you can use entirely on-device and offline with the Mac version. I havenât checked, but I assume that Google does have access to anything that gets processed in its cloud.
YeahâI donât think Iâd reach for it to transcribe an audio file if I had another tool to hand. Itâs good enough for the modest amount of uncomplicated dictation I do, and I like that itâs 100% local and private.
Iâm really glad. The transcription app I have been using ended up being clearly vibe coded as it is full of bugs and had a really badly optimized UI (not mentioning names, but this app is now a subscription, while it should never have been released - like most vibe coded apps!)
Iâm happy that this will wipe out a whole swath of recent vibe coded junk apps, but I feel sorry for more established companies in transcription who will likely not have a business going forward.
Not everyone wants to load a local model onto their computer.
People who do load local models onto their computers arenât likely a good target for a Gemma 4 only tool.
Not everyone has a computer that can run Eloquent as a daily driver.
Many users are willing to trade away the privacy upside of a local model for the power a cloud model offers. Eloquent doesnât have feature parity with, say, Whispr Flow.
Eloquent (for now) is English only.
I agree that Eloquent is a threat to vibe-coded slop, but I donât think itâs a threat to substantial, fully-featured transcription applications.
