Construct an EncoderLayer object.

Apply attention rescoring decoding, CTC prefix beam search is applied first to get nbest, then we resoring the nbest on atten

Export interface for c++ call, apply linear transform and log softmax before ctc Args: xs (torch.Tensor): encoder out

Export interface for c++ call, apply linear transform and log softmax before ctc Args: xs (torch.Tensor): encoder out

Apply CTC greedy search Args: speech (torch.Tensor): (batch, max_len, feat_dim) speech_length (torch.Tensor): (batch

Apply CTC prefix beam search Args: speech (torch.Tensor): (batch, max_len, feat_dim) speech_length (torch.Tensor): (

Export interface for c++ call, return eos symbol id of the model

Export interface for c++ call, return eos symbol id of the model

Apply forward pass.

Forward function. Args: xs: input tensor (B, L, D) Returns: output tensor, (B, L, D)

Forward decoder. Args: memory: encoded memory, float32 (batch, maxlen_in, feat) memory_mask: encoder memory mask, (ba

Forward decoder. Args: memory: encoded memory, float32 (batch, maxlen_in, feat) memory_mask: encoder memory mask, (ba

Compute decoded features. Args: tgt (torch.Tensor): Input tensor (#batch, maxlen_out, size). tgt_mask (torch.Tensor):

Frontend + Encoder + Decoder + Calc loss Args: speech: (Batch, Length, ...) speech_lengths: (Batch, ) tex

Forward decoder. Args: memory: encoded memory, float32 (batch, maxlen_in, feat) memory_mask: encoder memory mask, (ba

Forward decoder. Args: memory: encoded memory, float32 (batch, maxlen_in, feat) memory_mask: encoder memory mask, (ba

Input x. Args: x (torch.Tensor): Input tensor (#batch, time, idim). x_mask (torch.Tensor): Input mask (#batch, 1, tim

Subsample x. Args: x (torch.Tensor): Input tensor (#batch, time, idim). x_mask (torch.Tensor): Input mask (#batch, 1,

Subsample x. Args: x (torch.Tensor): Input tensor (#batch, time, idim). x_mask (torch.Tensor): Input mask (#batch, 1,

Subsample x. Args: x (torch.Tensor): Input tensor (#batch, time, idim). x_mask (torch.Tensor): Input mask (#batch, 1,

Frontend + Encoder + Decoder + Calc loss Args: speech: (Batch, Length, ...) speech_lengths: (Batch, ) tex

Compute loss between x and target. The model outputs and data labels tensors are flatten to (batch*seqlen, class) shape and a mask is

Return Swish activation function.

Calculate CTC loss. Args: hs_pad: batch of padded hidden state sequences (B, Tmax, D) hlens: batch of lengths of hidd

Args: x (torch.Tensor): (batch, max_len, feat_dim) Returns: (torch.Tensor): normalized feature

Compute scaled dot product attention. Args: query (torch.Tensor): Query tensor (#batch, time1, size). key (torch.Tens

Compute 'Scaled Dot Product Attention' with rel. positional encoding. Args: query (torch.Tensor): Query tensor (#batch, time1, siz

Compute convolution module. Args: x (torch.Tensor): Input tensor (#batch, time, channels). mask_pad (torch.Tensor): us

Embed positions in tensor. Args: xs: padded input tensor (B, T, D) xs_lens: input length (B) decoding_chu

Embed positions in tensor. Args: xs: padded input tensor (B, T, D) xs_lens: input length (B) decoding_chu

Add positional encoding. Args: x (torch.Tensor): Input. Its shape is (batch, time, ...) offset (int): position offset

Compute positional encoding. Args: x (torch.Tensor): Input tensor (batch, time, `*`). Returns: torch.Tensor: E

Just return zero vector for interface compatibility

Compute encoded features. Args: x (torch.Tensor): Input tensor (#batch, time, size). mask (torch.Tensor): Mask tensor

Compute encoded features. Args: x (torch.Tensor): (#batch, time, size) mask (torch.Tensor): Mask tensor for the input

Export interface for c++ call, forward decoder with multiple hypothesis from ctc prefix beam search and one encoder output Args:

Export interface for c++ call, forward decoder with multiple hypothesis from ctc prefix beam search and one encoder output Args:

Forward input chunk by chunk with chunk_size like a streaming fashion Here we should pay special attention to computation cache

Export interface for c++ call, give input chunk xs, and return output from time 0 to current chunk. Args: xs (torch.

Export interface for c++ call, give input chunk xs, and return output from time 0 to current chunk. Args: xs (torch.

Forward one step. This is only used for decoding. Args: memory: encoded memory, float32 (batch, maxlen_in, feat)

Forward one step. This is only used for decoding. Args: memory: encoded memory, float32 (batch, maxlen_in, feat)

Returns: torch.Tensor: decoder output

Returns: torch.Tensor: decoder output

Make mask tensor containing indices of non-padded part. The sequences in a batch may have different lengths. To enable batch computing, paddi

For getting encoding in a streaming fashion Attention!!!!! we apply dropout only once at the whole utterance level in a none

Apply beam search on attention decoder Args: speech (torch.Tensor): (batch, max_len, feat_dim) speech_length (torch.

Compute relative positinal encoding. Args: x (torch.Tensor): Input tensor (batch, time, size). zero_triu (bool): If tr

Following works like `RepVGG: Making VGG-style ConvNets Great Again` - https://arxiv.org/pdf/2101.03697.pdf. We re-parameterize multi-branche

Export interface for c++ call, return right_context of the model

Export interface for c++ call, return right_context of the model

Export interface for c++ call, return sos symbol id of the model

Export interface for c++ call, return sos symbol id of the model

Export interface for c++ call, return subsampling_rate of the model

Export interface for c++ call, return subsampling_rate of the model