Software is too slow for this task, especially if that speaking doll does not have a powerful CPU. It is way cheaper to produce a lot of cheap ASICs which can recognize very fast only the "ok google" keyword, maybe few most common phonemes also, but otherwise it just throws recorded sound on the server. So, all the software for recognizing unusual phrases is usually running on cloud and the speaking doll has nothing to update.
Its not baked into chip, its software and gets updated just like your phone or computer when its connected to the internet.