Teacher forcing is usually used, in which the decoder history is forced to be the correct gold $y_i$ rather than the predicted $\hat{y}_i$. It is a possibility to use a mixture of gold and decoder output. For example, one can use the gold output 90% of the time, but with probability .1 take the decoder output instead.

Columbia University

Encoder-decoders for speech are trained with normal cross-entropy loss, which is typically used for conditional language models. At timestep $i$ of decoding, the loss is the log probability of the correct token (which is a letter) $y_i$. 

Learning for LAS

An on-going but a helpful book resource about NLP
https://web.stanford.edu/~jurafsky/slp3/

Learn Before

Related