DeepSpeech is too large to run as a browser extension.
Have they tried nanonets?
Googled those, and it looks like some company selling remotely-running models, which I think is what GP was referring to. Is there another technique that's been SEOd out by this company?
They appear to be running it against both Google and their own DeepSpeech instance then sending both results back in the client's telemetry?