* @class tensor * @brief A TensorFlow eager tensor wrapper * */

* @class tensor * @brief A TensorFlow eager tensor wrapper * */

positional encoding can be used in transformer

Class to hold a set of hyperparameters as name-value pairs. A `HParams` object holds hyperparameters used to build and train a model, such as

TransformerEncoder is a stack of N encoder layers Args: encoder_layer: an instance of the TransformerEncoderLayer() class (required).

TransformerEncoderLayer is made up of self-attn and feedforward network. Args: d_model: the number of expected features in the input (req

The main text featurizer interface

Multi-head attention consists of four parts: * Linear layers and split into heads. * Scaled dot-product attention. * Concaten

Instance normalization layer. References - [Instance Normalization: The Missing Ingredient for Fast Stylization] (https://arxiv.or

Multi-head attention consists of four parts: * Linear layers and split into heads. * Scaled dot-product attention. * Concaten

Implementation of specaugument from paper "SpecAugment: A Simple Data Augmentation Method for Automatic Speech Recognition" Args: pre

Interface of audio features extractions. The kernels of features are based on Kaldi (Povey D, Ghoshal A, Boulianne G, et al. The Kaldi speech reco

Queue for data prefetching

Tensorflow ReflectionPad1d module.

Do CMVN on features. Args: config: contains four optional parameters. Shape: - output: :math:`(T, F)`. Examples::

Convolution Module. See section 2.2 of this paper for a description of this technique: Conformer: Convolution-augmented Transformer for Speec

Hold and save best metric checkpoint Args: name: MetricChecker name maximum: more greater more better

scaled positional encoding, reference: https://arxiv.org/pdf/1809.08895.pdf

Seq2SeqSparseCategoricalAccuracy Inherits CharactorAccuracy and implements Attention accuracy calculation

Seq2SeqSparseCategoricalCrossentropy LOSS CategoricalCrossentropy calculated at each character for each sequence in a batch

A transformer model. User is able to modify the attributes as needed. Args: d_model: the number of expected features in the encoder/decod

CTCAccuracy Inherits CharactorAccuracy and implements CTC accuracy calculation

CTC LOSS CTC LOSS implemented with Tensorflow

CharactorAccuracy Base class for Word Error Rate calculation

Convolutional 2D subsampling (to 1/4 length). Args: idim (int): Input dimension. odim (int): Output dimension. dropout_ra

positional encoding can be used in transformer U2

Wrapper for LSTM cell to create ZoneOut Cell inspired by: https://github.com/teganmaharaj/zoneout/blob/master/zoneout_tensorflow.py Publi

A wrapper for Tensorflow checkpoint Args: checkpoint_directory: the directory for checkpoint summary_directory: the directory for

A conformer model. User is able to modify the attributes as needed. The architecture is based on the paper "Attention Is All You Need". Ashish Vas

TransformerEncoder is a stack of N encoder layers Args: encoder_layer: an instance of the TransformerEncoderLayer() class (required).

TransformerEncoderLayer is made up of self-attn and feedforward network. This standard encoder layer is based on the paper "Attention Is All You N

Calculate duration and outputs based on teacher model

Fastspeech2 Feature Normalizer

Computing filter banks is applying triangular filters on a Mel-scale to the power spectrum to extract frequency bands. Args: confi

Feature Normalizer

python implementation of griffinlim algorithm

Compute spectrum features of every frame in speech. Args: config: contains ten optional parameters. Shape: - output: :mat

Tacotron2 Loss

Tensorflow BN_Residual_Block module.

CIN Block.

A conformer CTC model. User is able to modify the attributes as needed. The architecture is based on the paper "Attention Is All You Need". Ashish

TransformerDecoder is a stack of N decoder layers Args: decoder_layer: an instance of the TransformerDecoderLayer() class (required).

ConformerDecoderLayer is made up of self-attn, multi-head-attn and feedforward network. This standard decoder layer is based on the paper "Attenti

ConformerEncoder is a stack of N encoder layers Args: encoder_layer: an instance of the TransformerEncoderLayer() class (required).

ConformerEncoderLayer is made up of self-attn and feedforward network. This standard encoder layer is based on the paper "Attention Is All You Nee

A transformer model. User is able to modify the attributes as needed. Args: d_model: the number of expected features in the encoder/decod

TransformerDecoder is a stack of N decoder layers Args: decoder_layer: an instance of the TransformerDecoderLayer() class (required).

ConformerDecoderLayer is made up of self-attn, multi-head-attn and feedforward network. This standard decoder layer is based on the paper "Attenti

TransformerEncoder is a stack of N encoder layers Args: encoder_layer: an instance of the TransformerEncoderLayer() class (required).

TransformerEncoderLayer is made up of self-attn and feedforward network. This standard encoder layer is based on the paper "Attention Is All You N

Convolution Module. See section 2.2 of this paper for a description of this technique: Conformer: Convolution-augmented Transformer for Speec

Calculate duration and outputs based on teacher model

ExponentialDecayLearningRateSchedule Example: >>> optimizer = tf.keras.optimizers.Adam(learning_rate = ExponentialDecayLearningRate(0.01

Refered https://github.com/TensorSpeech/TensorFlowTTS

used for training of fastspeech2

used for training of fastspeech

Compute power of every frame in speech. Args: config: contains four optional parameters. Shape: - output: :math:`(T, 1)`.

GuidedAttention Loss to make attention alignments more monotonic

Guided multihead attention loss function module for multi head attention.

Length regulator for Fastspeech.

location-aware attention Reference: Attention-Based Models for Speech Recognition (https://arxiv.org/pdf/1506.07503.pdf)

MPC LOSS L1 loss for each masked acoustic features in a batch

Compute pitch features of every frame in speech. Args: config: contains nineteen optional parameters. Shape: - output: :mat

Calculate the attention weights. q, k, v must have matching leading dimensions. k, v must have matching penultimate dimension, i.e.: seq_len_k

Calculate the attention weights. q, k, v must have matching leading dimensions. k, v must have matching penultimate dimension, i.e.: seq_len_k

SpeechSynthesisDatasetBuilder

Dataset builder for RNN model. The builder mix the spliced frame in one dim For example (1, 1323) The input data format is (batch, t, dim

Dataset builder for CNN model. The builder treat every spliced frame as one image. For example (21, 63) The input data format is (batch,

Stepwise monotonic attention Reference: Robust Sequence-to-Sequence Acoustic Modeling with Stepwise Monotonic Attention for Neural TTS (https://a

TextTokenizer

TransformerDecoder is a stack of N decoder layers Args: decoder_layer: an instance of the TransformerDecoderLayer() class (required).

TransformerDecoderLayer is made up of self-attn, multi-head-attn and feedforward network. Reference: "Attention Is All You Need".

A transformer model. User is able to modify the attributes as needed. Args: d_model: the number of expected features in the encoder/decod

TransformerDecoder is a stack of N decoder layers Args: decoder_layer: an instance of the TransformerDecoderLayer() class (required).

TransformerDecoderLayer is made up of self-attn, multi-head-attn and feedforward network. Reference: "Attention Is All You Need". Ar

TransformerEncoder is a stack of N encoder layers Args: encoder_layer: an instance of the TransformerEncoderLayer() class (required).

TransformerEncoderLayer is made up of self-attn and feedforward network. Args: d_model: the number of expected features in the input (req

Vocabulary

WarmUp Learning rate schedule for Adam Example: >>> optimizer = tf.keras.optimizers.Adam(learning_rate = WarmUpLearningSchedule(512),

WarmUp Learning rate schedule for Adam and can initialize a learning rate Example: >>> optimizer = tf.keras.optimizers.Adam(learning_rat

Additive Angular Margin Softmax Loss Reference to paper "ArcFace: Additive Angular Margin Loss for Deep Face Recognition"

abstract of Frontend

Additive Margin Softmax Loss Reference to paper "CosFace: Large Margin Cosine Loss for Deep Face Recognition" and

DecoderSolver

A multi-processer solver based on Horovod

Base Solver.