Clones the current state while overwriting components provided by kwargs.

Initializes the model for inference sets "mel_outputs" and "alignments" fields. Args: - inputs: int32 Tensor with

Mu-Law companding Method described in paper [1]_. .. math:: f(x) = sign(x) ln (1 + mu |x|) / ln (1 + mu) Args: x (array-like): Input signal. Ea

Mu-Law companding + quantize Args: x (array-like): Input signal. Each value of input signal must be in range of [-1, 1]. mu (number): Compres

conv1x1 step

Initialize wavenet graph for train, eval and test cases.

Inverse of mu-law companding (mu-law expansion) .. math:: f^{-1}(x) = sign(y) (1 / mu) (1 + mu)^{|y|} - 1) Args: y (array-like): Compressed sign

Save Tacotron-2 training state to disk. (To skip for future runs)

Returns a 2-D or 3-D tensorflow sequence mask depending on the argument 'expand'

Converts a string of text to a sequence of IDs corresponding to the symbols in the text. The text can optionally have ARPAbet sequences enclose

Computes a masked Mean Squared Error

Expand global conditioning features to all time steps Args: batch_size: int time_length: int global_features: Tensor of shape [batch_size, cha

returns the last word in a string as a float

Returns all-zero <GO> frames for a given batch size and output dimension

librosa implementation of Griffin-Lim Based on https://github.com/librosa/librosa/issues/434

Adds loss computation to the graph. Supposes that initialize function has already been called.

Adds loss to the model. Sets "loss" field. initialize must have been called.

Build `Layer`

Load shadow variables of saved model. Inspired by: https://www.tensorflow.org/api_docs/python/tf/train/ExponentialMovingAverage Can also use: shad

Inceremental forward step Inputs of shape [batch_size, channels, time_length] are reshaped to [batch_size, time_length, channels] Input of each t

Converts mel spectrogram to waveform using librosa

Inverse of mu-law companding + quantize Args: y (array-like): Quantized signal (∈ [0, mu]). mu (number): Compression parameter ``μ``. Returns:

numerically stable log_sum_exp implementation that prevents overflow

lowercase input tokens.

Args: y: Tensor, [batch_size, channels, time_length] Returns: Tensor: sample in range of [-1, 1]

sample from learned gaussian distribution

A function that splits any string based on specific character (returning it with the string), with maximum number of words on it

Args: x: Tensor [batch_size, channels, time_length] c: Tensor [batch_size, c_channels, time_length]. Local conditioning features g: Tenso

Return a generator of HTS format label of txt. Args: txt: like raw txt "向香港特别行政区同胞澳门台湾同胞" or txt with prosody make like "向#1

Computes a masked MAE loss with priority to low frequencies

Computes a masked SigmoidCrossEntropy with logits

Initialize encoder parameters Args: convolutional_layers: Encoder convolutional block class lstm_layer: encoder bidir