Yea, I was wondering how they dealt with word boundaries and different word audio lengths generally (short and long). Your point about parallelism is interesting. Doing what they say they are doing, this should be ridiculously parallel, possibly with a reduce step at the end to remove redundant words if their slicing algorithm allows for that to happen. But yea, extremely parallel.