Wrapper class of torch.nn.Linear Weight initialize by xavier initialization and bias initialize to zeros.

Residual Connection Module. outputs = (module(inputs) x module_factor + inputs x input_factor)

Conformer encoder first processes the input with a convolution subsampling layer and then with a number of conformer blocks. Args:

input : (BxLxD) output : (BxLxD)

Conformer Feed Forward Module follow pre-norm residual units and apply layer normalization within the residual unit and on the input before t

Create a fbank from a raw audio signal. This matches the input/output of Kaldi’s compute-fbank-feats Args: sample_rate (int): Samp

When kernel size == 1 conv1d, this operation is termed in literature as pointwise convolution. This operation often used to match dimensions.

Swish is a smooth, non-monotonic function that consistently matches or outperforms ReLU on deep networks applied to a variety of challenging

Conformer block contains two Feed Forward modules sandwiching the Multi-Headed Self-Attention module and the Convolution module. This sandwic

Conformer convolution module starts with a pointwise convolution and a gated linear unit (GLU). This is followed by a single 1-D depthwise co

Convolutional 2D subsampling (to 1/4 length) Args: in_channels (int): Number of channels in the input image out_channels (in

Decoder of RNN-Transducer Args: num_classes (int): number of classification hidden_state_dim (int, optional): hidden state d

When groups == in_channels and out_channels == K * in_channels, where K is a positive integer, this operation is termed in literature as dept

The gating mechanism is called Gated Linear Units (GLU), which was first introduced for natural language processing in the paper “Language Mo

Inputs : outputs : tuple ( tensor (BxSxE), tensor (BxLxE) ) targets : tensor (BxS)

Inputs : (B x S x E), (B x T x E)

Conformer employ multi-headed self-attention (MHSA) while integrating an important technique from Transformer-XL, the relative sinusoidal pos

Noam learning rate scheduler with piecewise linear increase and exponential decay. The learning rate increases linearly from init_lr to max

Positional Encoding proposed in "Attention Is All You Need". Since transformer contains no recurrence and no convolution, in order for the mo

Multi-head attention with relative positional encoding. This concept was proposed in the "Transformer-XL: Attentive Language Models Beyond a

input : (BxCxL) output : (BxDxL`) L`:= length of audio sequence with 30 Hz

input : (BxCxLxHxW) output : (BxLxD)

Computes the Word Error Rate, defined as the edit distance between the two provided sentences after tokenizing to words.

Inputs : (B x S x E), (B x T x E)

Wrapper class of torch.transpose() for Sequential module.

Computes the Word Error Rate, defined as the edit distance between the two provided sentences after tokenizing to words.

Conformer: Convolution-augmented Transformer for Speech Recognition The paper used a one-lstm Transducer decoder, currently still only implem

Provides inteface of error rate calcuation. Note: Do not use this class directly, use one of the sub classes.

Create the Mel-frequency cepstrum coefficients (MFCCs) from an audio signal. Args: sample_rate (int): Sample rate of audio signal.

Create a spectrogram from a audio signal. Args: sample_rate (int): Sample rate of audio signal. (Default: 16000) frame_le

Wrapper class of torch.view() for Sequential module.

Note: Do not use this class directly, use one of the sub classes.

Note: Do not use this class directly, use one of the sub classes.