search

Rayhane-mamah / Tacotron-2

DeepMind's Tacotron-2 Tensorflow implementation
MIT License
2.27k stars 905 forks source link

Location Sensitive Attention vs BahdanauMonotonic Attention #338

Open hccho2 opened 5 years ago

hccho2 commented 5 years ago

Location Sensitive Attention(step 8500) test-step-000008500-locationsensitive-attention

BahdanauMonotonic Attention(step 8500) test-step-000008500-bahdanau-attention

These plots show that BahdanauMonotonic Attention is better.

What are the advantages of Location Sensitive Attention?

begeekmyfriend commented 5 years ago

Show us the samples please? By the way, you had better change the mel loss function into MAE and watch the alignment again.

hccho2 commented 5 years ago

The results made by location sensitive attention are a little worse(in my opinion).

  1. Do not the alignment plot and the results(audio quality) match?
  2. What are the advantages of Location Sensitive Attention?
begeekmyfriend commented 5 years ago

As for Chinese mandarin, the evaluation results sound good for me. Here is my T2 fork and https://github.com/Rayhane-mamah/Tacotron-2/issues/292#issuecomment-444823633 is my samples.

begeekmyfriend commented 5 years ago

Did the synthesized results have any misreading or word missing based on Bahdanau Monotonic Attention compared with Location Sensitive Attention?

hccho2 commented 5 years ago

I can not be sure because I am under training.

The following plot with Location Sensitive Attention shows missing words. test-step-000024000-align001

linlinsongyun commented 5 years ago

As for Chinese mandarin, the evaluation results sound good for me. Here is my T2 fork and #292 (comment) is my samples.

which is better?location sensitive attention?