There's an interesting contrast between that slick faked demo with the seemingly live video recognition and another video Google released the same day, showing the processing and understanding of raw audio [1]. In the latter, the awkward interface used in the demonstrations looks like something cobbled together for internal use, and the first example, of recognition of tones in spoken Mandarin, looks like something from a research paper, not a consumer-facing product announcement.
That said, the use of LLMs for language learning is one of my main interests, and the ability to detect nuances in a learner's pronunciation will be very useful when it becomes widely available.
[1] https://youtu.be/D64QD7Swr3s