# Speech Recognision

## I/O of SR sys

## Feature Extraction before entering NN

## Acoustic features used in paper

## How much data needed

## Famous Structure

## Audio signal NN pooling

## Listen, Atten, and Spell (LAS)
Inference: Beam Search
- Greedy Search and its limit
- Beam Search

Training: Teacher forcing

Attention


## Connectionist Temporal Classification(CTC)
* Input T acoustic features, output T tokens (ignoring down sampling)
* Output tokens including $\phi$, merging duplicate tokens, removing$\phi$

## Model summary