2018.09.17 Tacotron 2 + Wavenet
Speaker 0 (Regina)
Initial problems with:
- Embedding settings for wavenet. Missed setting gin_channels parameter, which defines embedding layer size (for multiple voices).
- Dataset preprocessed with different parameters. Wavenet with fmin 125 and taco with fmin 0, which caused voice pitch shift.
Text lines
- Sveiki, draugai
- Aš robotas iš ateities
- Aš žinau jūsų likimą
Gin 16, 130000 Step
-
-
-
Gin 16, 270000 Step, taco 100000 step
-
-
-
Gin 16, wavenet 2780000 Step, taco 110000 step
-
-
-
No wavenet embedding
-
-
-