Description

I trained Librispeech problem days ago and follow the notebook demo in https://github.com/tensorflow/tensor2tensor/blob/master/docs/tutorials/asr_with_transformer.md
when decoding. The training data is 100h+360h and the model I used for decoding has the following status:

INFO:tensorflow:Finished evaluation at 2018-05-25-13:32:33 INFO:tensorflow:Saving dict for global step 120002: global_step = 120002, loss = 1.0433111, metrics-librispeech/targets/accuracy = 0.633423, metrics-librispeech/targets/accuracy_per_sequence = 0.0, metrics-librispeech/targets/accuracy_top5 = 0.8924546, metrics-librispeech/targets/edit_distance = 0.5968527, metrics-librispeech/targets/neg_log_perplexity = -1.2044775 INFO:tensorflow:Stop training model as max steps reached FLAGS.hparams_set:: transformer_librispeech FLAGS.hparams:: batch_size=4096,

I decoded the test data(those files are generated by t2t-datagen together with the training data), however the results are not quite as I expect
decode_from_file/61-70968-0001.flac.wav
Outputs: AND THAT THEY WERE TO BE SEEN THEY WERE ALL THE SAME AS THEY WERE TO BE ABLE TO SEE THEM
decode_from_file/61-70968-0002.flac.wav
Outputs: AND THAT THEY WERE ALL THE SAME AS THEY WERE
decode_from_file/61-70968-0003.flac.wav
Outputs: AND THAT THEY WERE ALL THE SAME
decode_from_file/61-70968-0004.flac.wav
Outputs: AND THAT THEY WERE ALL THE SAME AS THEY WERE ALL THE SAME

while the answers should be:
61-70968-0001 GIVE NOT SO EARNEST A MIND TO THESE MUMMERIES CHILD
61-70968-0002 A GOLDEN FORTUNE AND A HAPPY LIFE
61-70968-0003 HE WAS LIKE UNTO MY FATHER IN A WAY AND YET WAS NOT MY FATHER
61-70968-0004 ALSO THERE WAS A STRIPLING PAGE WHO TURNED INTO A MAID

Question

1 is the model properly trained?
2 what could be wrong if I want a decent result?