One voice detector to rule them all

We have just updated our public voice detector, now it boasts features like:

- Epic metrics (best in class as far as we know); - Much shorter and flexible chunk size (30, 60, 100 ms); - Radically faster inference, i.e. model forward pass takes ~1ms for a 30 - 100ms chunk; - Radical simplification - there is literally just one model for everything now;