sciforce / phones-las Goto Github PK

Hi,

I am doing some ASR experiments using AFs and I found your work, which is fantastic work.

And I noticed that you mentioned this framework can extend to cross-lingual ASR, Did you already have some papers?

Meanwhile, I am trying to replicate your work, hope it goes well

Thanks in advance

Update README.md with info on how to publish results to tensorboard.dev.

With all the new things since TF 2.0 release, it may be reasonable to port code to newest version.
Another option is migration to PyTorch. :)

In preprocess_all.py:

            text = get_ipa(text, language, remove_all_diacritics=args.remove_diacritics,
                split_all_diphthongs=args.split_diphthongs)

This raises a huge amount of errors in output, because _postprocess from ipa_utils.py doesn't have remove_all_diacritics param.

Hi sciforce! This repo looks really nice.

I'm having trouble running this code with TIMIT and, in particular, with configuring the espeak-ng dependency.

Could you add a line to the readme on installing espeak-ng? Does this dependency even support mac?

First of all, I would like to thank you for sharing this great work.

I trained the model successfully based on the instructions in README for articulatory detection in TIMIT.

But, I faced some problems in infer.py.

1. This is the type of error "operands could not be broadcast together with shapes (62,) (42,) " in this function "segs_phones_to_frame_binf". Because there is 62 attributes and 42 phonemes [39 (+UNK, SOS, EOS)]. The length of logits_binf is only 42.

2. Also, could you provide us the Notebook for compute the accuracy of attribute detection at frame level.

3. Finally, in the paper, there are 28 place and manner attributes. My question, which mapping file that you used to calculate the detection accuracy of these 28 only.

Thanks a lot.