Open guangsenw opened 5 years ago
Thanks! We didn't try out TIMIT, but tried out the toy dataset, as well as RIMES handwriting (as these resemble audio signals). We found that the results were not significantly better.
The code is quite old now and is quite hacky at times. There is an implementation in RETURNN: https://github.com/rwth-i6/returnn with a demo here: https://github.com/rwth-i6/returnn/blob/master/demos/demo-tf-neural-transducer.12ax.config The implementation there isn't exactly as in the paper, as we don't go "blockwise" over the input.
The almost exact implementation can be seen in this repo, i suggest looking at the managing script: https://github.com/nikita68/i6-seq2seq/blob/master/neural-transducer/nt_rimes.py
Nice work. Just wondering whether the implementation can be used to reproduce the TIMIT result reported in the paper. If not TIMIT, what datasets have been used to test the code and are there some known issues or work to be done with the code? thanks