Table 5 The hyperparameters of the alignment and synthesis model

Batch size

16

Dropout

0.2

Adam optimizer \(\beta _1\)

0.9

Adam optimizer \(\beta _2\)

0.98

Adam optimizer \(\epsilon\)

\(10^{-9}\)

Learning rate annealing steps (alignment)

1000

Learning rate annealing rate (alignment)

0.5

Gradient clipping threshold

1.0

Warm up steps (alignment)

1000

Warm up steps (synthesis)

4000

Training steps (alignment)

10,000

Training steps (synthesis)

100,000