PostProcessing Config via Transformers
NeMo-Speech.cpp and the Python nemo-toolkit offer parameters for configuring post-processing (e.g. min_duration_on, min_duration_off). However, I don't currently see a way to configure those settings through the Transformers API. Is that in the works?
As reference, on my small dataset, I saw a DER of 15.37% with transformers, but a DER of 5.69% with NeMo-Speech.cpp. The model hashes are the same for those cases, so I chalk it up to pre-/post-processing.
My current workaround is to build my own Python wrapper around Nemo-Speech.cpp, which seems to be frequently updated. The nemo-toolkit is nice, but includes many training dependencies I don't need.
But I already have a dependency on transformers, so if there were a way to incorporate post-processing in that API, I would love it.
Hey !! thanks for requesting this feature, let me raise a PR to add it!