levhaikin

I ran into the same issue. **tl;dr** - the following line is a possible cause of the issue (but it might not be the ultimate root-cause): https://github.com/parlance/ctcdecode/blob/9a20e00f34d8f605f4a8501cc42b1a53231f1597/ctcdecode/src/path_trie.cpp#L44 this line allows...

@ShantanuNair not sure it's related directly.. I have never tried feeding the network and beam-decoder with such long sequences.. a common practice is to break the long audio into distinct,...

hi, I ended up reimplementing the whole algorithm from scratch in c++.. came to the conclusion that merging paths that have same "squeezed" form loses information that can't be recovered...

@panmareksadowski regarding efficiency, I didn't run back-to-back tests comparing the two codebases, but generally they seem to be in the same ballpark, at least concerning speed. I can say that...