Juntae, Kim

Acc. to line 181 in ram.py the c.e loss and location policy*reward were summed and the summation operator is just gradient router when backpropa phase, i.e. c.e. part and location...

Hi, If you want to increase ASR performance using enhanced data, you should train your AM model with enhanced dataset with noisy dataset.

noisy_speech = v_addnoise(speech, desired_fs, snr_list(snr_indx), 'k', noise, noise_fs); the noise_fs is input sampling rate and desired_fs is the output sampling rate, therefore, it seems no problem.

p{margin-top:0;margin-bottom:0} You can find that function from  http://www.ee.ic.ac.uk/hp/staff/dmb/voicebox/voicebox.html Thx.  Jaytee.KimSpeech Recognition Part / Research Engineer / Ph. [email protected] Corp.8F, 235, Pangyoyeok-ro, Bundang-gu,Seongnam-si, Gyeonggi-do, 13494, Korea --------- 원본 메일 --------- 보낸사람:...

I'm very sorry for late answer the pre-trained models were already uploaded in backup folder thx.

1. The problem occurs from the mismatch of datareader and model's placeholder. According to your debugging, the datareader outputs the dnn's inputs but the model is set to lstm. Please...

Also the first problem you mentioned is not the code problem, but some environmental problem. How about changing your environment to ubuntu 16.04 from window? Actually, all of my projects...

Don't change the code and at first, just set mode='fnn' (basic mode) and in main.py try to set reset=True, train_reset=True, test_only=False Also, clean your files in logs. and remove some...

NO. this toolkit does not need vad code or that pb file. That code is just legacy code which is not necessary.

Thank you for your reply. I figure out that that error is from version difference of TensorboardX. I'll fix that part, soon.