[toc]

Models

Dataset

Token

utterance，发音的最小单元

phoneme，音素，和lexicon一起使用

grapheme

word,

morpheme

bytes，直接用utf-8编码，language independent

Feature Extraction

Down Sampling

Pyramid RNN

TDNN

实质是CNN

限制self-attention的距离

Attention

z0是关键字，α由z0和ht得到

RNN-T with language model