[NLP Li Hongyi] notes

One , Six models

Input a sound ------> The output text 、 voice 、 Judge category

 

Enter the language judgment category ：

 

Two 、speech recognition Voice recognition

Three major speech recognition conferences , Use the most token：

  Speech recognition +xxxx：

  Sound features ：

 

 

  Species speech recognition model

        LAS（Listen,Attend,and Spell ）

 

 encoder May adopt rnn,cnn

 

  1, obtain attention Two ways ：

z0 Keywords and every database （encoder Output ）h1、h2.... Enter a name respectively match Of function Work out a skeler That is, a value .

 

frequently-used match  function The way ：

 2,spell