Ruchao Fan
Ruchao Fan
不使用语言模型beam search还是能提升性能,因为每条path中blank的位置会影响最后的解码结果。
没有在这个repo下试过,猜测multi-head attention应该需要调整一下学习策略。
@iamxiaoyubei The best PER of TIMIT is 16.9% and almost 11% CER for partial 863 Chinese corpus without careful training. Maybe I will re-run the repo and update an averaged...
@iamxiaoyubei Sorry to tell you that there are no results on Librispeech at present. As I mentioned, maybe it will be reported in the future.
Seems not work for me with cuda9.2 and V100. With "./test_cpu: error while loading shared libraries: libwarpctc.so: cannot open shared object file: No such file or directory". Any suggestion?