"this can be very easily done fully client-side" : maybe, if you have the voice model and and inference engine that runs well on devices. Mozilla doesn't have t

"this can be very easily done fully client-side" : maybe, if you have the voice model and and inference engine that runs well on devices. Mozilla doesn't have that yet, so this experiment uses a backend running a Kaldi server and model that uses too much memory to run locally. Once DeepSpeech is ready I'm pretty sure they will switch to that, and ultimately to on-device voice recognition with PipSqueak (PipSqueak is expected to be an inference engine usable on devices). Unfortunately none of these projects are far along enough to be usable. Common Voice is mostly related to DeepSpeech as this will help getting data to train the engine.