Xiao
Xiao
It's because of the ctc-loss is not well integrated in Keras. To use ctc loss I have to generate output by base model, and use the output as the input...
https://www.dropbox.com/s/anvw5x7bmjsm94u/listen_model.chk?dl=0
are you using the same version of python and keras? 2018-03-02 17:02 GMT+08:00 zhaowz : > use your code and load your model but have error, can you help me?...
> https://www.dropbox.com/s/ujiardfgl2z51ya/listen_model.chk?dl=0
> This was done too long ago, I forget the details. If you can find the Chinese audio dataset I used in this repositary, it will not be hard to...
y不起作用，ctc应该降为1以下才算收敛。
"mfcc_vec.npy" can be generated by the code. Since it's more than 500 MB, it's not convenient to upload here.
This is not the result I have got. There is something wrong with your code.
所有的code都上传了 没有其它技巧。
@pingchesu Indeed! However the quality of the dataset is very important. mfcc features needn't to be very large, unless enough GPU power is possessed.