• transforms the phonemes into a hidden

• 3-layer CNN bidirectional LSTM

• it is same as in Tacotron2

• input

  • phoneme $t$ (B, T)

• output

  • phoneme representation $h_{text}$ (B, 512, T)

• model architecture

  TextEncoder(
    (embedding): Embedding(178, 512)
    (cnn): ModuleList(
      (0-2): 3 x Sequential(
        (0): ParametrizedConv1d(
          512, 512, kernel_size=(5,), stride=(1,), padding=(2,)
          (parametrizations): ModuleDict(
            (weight): ParametrizationList(
              (0): _WeightNorm()
            )
          )
        )
        (1): LayerNorm()
        (2): LeakyReLU(negative_slope=0.2)
        (3): Dropout(p=0.2, inplace=False)
      )
    )
    (lstm): LSTM(512, 256, batch_first=True, bidirectional=True)
  )
  

Why not to use transformer like VITS?