Mycroft: an open-source voice assistant

Voice recognition via AI techniques has a number of "embarrassingly parallel" steps. So if you can offload the problem to a large number of servers then responses are timely. As the parameters of the STT problem are better understood then it becomes economic for GPU designers (who are increasingly "parallel processor" designers) to offer tuned hardware for local processing. At the moment research groups are getting good results from high-end NVIDIA GPUs, which have instruction sets for highly-parallel tensor processing.