At the DDD last week I was talking to a young lad who had been to my IoT talk and wanted to make a voice controlled device. At the time we talked I was thinking that this is the kind of thing that you would push into the cloud for processing. However, TensorFlow Lite is now a proper part of the Arduino libraries (you can just add it as you would any other device driver - search for Tensor).
One of the examples actually does voice recognition on the device. I’ve not found anyone who has got this working on the ESP32 but it would be wonderful if it was possible.