Problem when train with LSTM

[johnnyxuan]

I have trained slt_arctic_full demo using default configuration, and the result is fine. but when I changed the configuration as bellows:

[Architecture] hidden_layer_size : [1024, 1024, 1024, 1024, 1024, 512] hidden_layer_type : ['TANH', 'TANH', 'TANH', 'TANH', 'TANH', 'LSTM'] #if RNN or sequential training is used, please set sequential_training to True. sequential_training : True

The result is not correct as I expected, is there any problem with my configurations? The synthesized audio files are as shown in the attachment. Thanks.

arctic_lstm.zip

[malradhi]

This issue is still remain same !!! any suggestions?

[ZhaoZeqing]

@johnnyxuan Hi John, I also encountered this problem. How many steps when you trained this model? I guess the training_epochs and the warmup_epoch in the conf file need to be larger when using LSTM.