INTRODUCTION TO FINITE STATE MACHINES
Jie Zhao
Institute for Signal and Information Processing
Mississippi State University, Mississippi State, MS 39762
email: zhao@isip.msstate.edu

ABSTRACT

Finite state machine (FSM) is a network which has finite number of states. Given an input, the machine can transfer from one state to another and generate an output. In speech recognition, each acoustic model (usually the Hidden Markov model) and each language model is a small FSM. All these small FSMs compose a big and complicated FSM or network, that is the search space in decoding. The aim of decoding is to find the best path among this big network. Therefore understanding how FSMs operate is fundamental to the research on speech recognition.

In this talk, I will introduce the fundamental theories about FSM and its applications in speech recognition :

• Types of FSMs: deterministic and nondeterministic
  
  
• Operations on FSMs: Union (sum), Concatenation, Difference, Composition, Determinization, Best path, etc. And along with AT&T FSM tutorial, I'll talk about the algorithms to do these operations.
  
  
• FSMs in speech recognition
  
  
  • Hidden Markov Models
  • N-gram language model

Additional items of interest:

• Presentation
• Paper
• Software, Data, etc.