Hi all,
I have implemented the Wyoming protocol into the Whisper TPU project. It loads in the model and uses it everytime a request from Home Assistant comes in, speeding up the workflow.
Compared to FasterWhisper (CPU based) I would say the TPU implementation is 20-30% faster.
Test data shows that for Whisper CPU + Piper CPU, a command like “Turn on the living room lights” is fully computed within 5-6 seconds.
With Whisper TPU + Piper CPU, it takes 3-4 seconds, so a nice increase.
Project can be found here: https://github.com/RaspberryTech01/Wyoming-Whisper-TPU/tree/release