Google launches Gemma-powered dictation app; here’s the first look

Google quietly launched its dictation app—AI Edge Eloquent.

Apple users can download it from the App Store for free. What makes the app stand out is that it has a Cloud-off mode, which means that all chats and transcriptions stay on the device, using local or private processing for cleanup. In Cloud-on mode, the audio still remains on the device but uses Gemini models to clean up the transcripts. 

The app is powered by Gemma-based automatic speech recognition (ASR) models. Gemma is a family of lightweight, state-of-the-art open models built from the same research and technology used to create the Gemini models.

Here’s the first look:

The app launches after the models are fully loaded, which takes a few minutes, and opens into a simple screen with a record button and four options — Record, History, Dictionary, and Settings.

The Dictionary option allows you to sync jargon from your Gmail accounts, while also letting users add other words to improve transcript accuracy.

How does the Edge Eloquent perform?

One has to wait a few minutes for the app to launch properly; otherwise, it shows an alert that the models weren’t loaded to record and comprehend.

After the first trial, which was done in “Cloud off” mode, it calculated the total words spoken and cadence. Under the transcript, apart from “Polish” mode—which cleans up fillers like “um” and “uh”—it also deletes redundancy. It also shows the number of “Inserts” and “Deletes,” which are edits made to the original transcript.

To access other modes like Key Points, Short, Formal, and Long, one needs to be connected to the Cloud, as these require Gemini for processing.

Is this available for other platforms?

Edge Eloquent is available only on iOS for now. Google, in its latest blog post, said it is evaluating other platforms, including desktop, and use cases like dictating documents, code, and prompting AI agents.

The app is available only in English in most regions, except the UK, Switzerland, and the EEA due to regulatory approval requirements. Sometimes, the model may translate a few words from other languages like Hindi, but this feature is not supported yet, and multilingual models are expected in future releases.