Finally after some waiting time I got the TinyML board from Syntiant with the NDP 101 Processor in my hands. The examples provided worked out-of the box, understanding of stop and go is good, still having more false-positive than expected.
I am still not sure how well his network structure ( 3 Dense Layers of 256 neurons and 3 Dropout layers) and performance will compare against a 2 layer Conv1D (14/25) running on an ESP32 - that I used in a project to set-up an eInk Voice Controlled Alarm Clock. For sure, it will need less power.
I am summarizing my progress here: https://44-2.de/syntiant-ndp-101-always-on-low-power-speech-recognition/.
If you have some insights on the “always on voice-wakeup” set-up - let me know.
I will keep updating my blog - just dont expect it to be to regularly