Honestly, handling a realistic voice is incredibly hard; being able to parse context, including from structured, partial information, even more so. I don’t know

Honestly, handling a realistic voice is incredibly hard; being able to parse context, including from structured, partial information, even more so. I don’t know if the progress in training processors, LSTM structure in the next year, but I doubt many companies will have enough conversations to train their models -- other than the operator of, say, Google Voice. I am willing to believe that any company who can offer that service as a service will know that they need to monitor for abuse, and have a very sophisticated team to catch those. That might include Apple, Amazon, Baidu (who probably won’t launch it in English). Facebook wasn’t able to see something like that working and they abandoned M four months ago.