Args: layers_sizes: list of integers, the length of the list represents the number of pre-net layers and the list

Args: is_training: Boolean, determines if the model is in training or inference to control zoneout layers: integer, t

Args: shape: integer, dimensionality of output space (r*n_mels for decoder or n_mels for postnet) activation: callabl

Args: is_training: Boolean, to control the use of sigmoid function as it is useless to use it during training sin

Args: is_training: Boolean, determines if the model is training or in inference to control dropout kernel_size: tuple

Construct the Attention mechanism. Args: num_units: The depth of the query mechanism. memory: The memory to query; usually the output of an RN

txt2pinyin in one file

Preprocesses a single utterance wav/text pair this writes the mel scale spectogram to disk and return a tuple to write to the train.txt file Ar

Preprocesses a single utterance wav/text pair this writes the mel scale spectogram to disk and return a tuple to write to the train.txt file Ar

Applies a smoothing normalization function instead of softmax Introduced in: J. K. Chorowski, D. Bahdanau, D. Serdyuk, K. Cho, and Y. Ben- gio,

Basic pipeline that lowercases and collapses whitespace without transliteration.

Build `Layer`

Preprocesses the speech dataset from a gven input path to given output directories Args: - hparams: hyper parameters - input_dir: input direct

Preprocesses the speech dataset from a gven input path to given output directories Args: - hparams: hyper parameters - input_dir: input direct

Preprocesses the speech dataset from a gven input path to given output directories Args: - hparams: hyper parameters - input

Call `Layer`

Call 'Layer

Pipeline for English text, including number and abbreviation expansion.

At sequential inference times: we adopt fast wavenet convolution queues approach by saving pre-computed states for faster generation inputs: [bat

Initialize the decoder. Args: name: Name scope for any created operations. Returns: `(finished, first_inputs, initial_state)`.

Converts linear spectrogram to waveform using librosa

compute right padding (final frame)

numerically stable log_softmax implementation that prevents overflow

Returns list of ARPAbet pronunciations of the given word.

Applies layer normalization. Args: inputs: A tensor with 2 or more dimensions, where the first dimension has `batch_size`. ep

Compute left and right padding

The `state_size` property of `TacotronDecoderCell`. Returns: An `TacotronDecoderCell` tuple containing shapes used by this object.

Perform a custom decoding step. Enables for dyanmic <stop_token> prediction Args: time: scalar `int32` tensor. inputs: A (structure of) inpu

Pipeline for non-English text that transliterates to ASCII.

Trim leading and trailing silence Useful for M-AILABS dataset if we choose to trim the extra 0.5 silence at beginning and end.