Bad inference for really short sentences

Rayhane-mamah / Tacotron-2

DeepMind's Tacotron-2 Tensorflow implementation

MIT License

2.27k stars 905 forks source link

Bad inference for really short sentences #370

Closed alexdemartos closed 5 years ago

alexdemartos commented 5 years ago

Hi,

I am experiencing this problem for a long time with many different models I was able to train.

At inference time, if the sentence contains 3 or more words it will work correctly most of the times. On the other hand, if a sentence is 1-2 words long ("Hi", "Hello"), it will generate unintelligible speech 90% of the times.

Any idea what might be causing this?

GregCKrause commented 5 years ago

Hi Alex,

Were you able to find a solution?