Google launched its voice recorder last year, able to transcribe in real time on devices such as Google Pixels. As in virtually all Google apps, this is possible through machine learning, which allows apps to constantly evolve.
Good proof of this is that Google has now incorporated a new feature called Smart Scrolling, another feature based on machine learning. Thanks to this, the most important words of the transcription are automatically chosen, so that they then appear in the toolbar scroll vertical. This way we can scroll through these keywords or touch them to navigate faster.
Google will recognize the most important words in the transcript
Google just announced Smart Scrolling, a machine learning-based feature for its Recorder app. With this new functionality, the application will be able to choose the keywords you consider to be keywords in the record, to highlight them when we do scroll in the transcript.
Google guarantees that, being a lightweight model, it can work locally, without the need for a network connection.
The learning model, according to Google, is so light that can be run offline, just from the device. How this feature works is quite interesting, as they detailed in their blog.
This template consists of two main tasks: one extracts the keywords from each section of the transcript and the other chooses the most informative sections of the text. Thanks to different text extraction models, they can process all tasks in parallel, analyze the context of each word in the text.
But how can Google know if a word is relevant? Using a learning model, the number of times the word appears in the text, the specificity of the term, i.e. if this word has a large semantic field or if it is very specific and other data on said word. In this way, the model is able to know the relevance of the word in the text, in a very summarized and simplified way (all this is much more complex, since it is pure computing). Google has tested this feature on the Pixel 5 and has yet to give any clues as to when it will arrive on all devices.
Via | Google AI Blog
