Often speech recognition has some tunable parameters that influence
performance. It is impossible to give an exhaustive list of possible
recogniser-specific parameters. Typical parameters are:
energy threshold s for word detection, minimum time gap between words,
acceptance threshold , noise rejection and input gain.
Sometimes the character of the recogniser can be tuned, by setting
parameters such as the number of states in the Hidden Markov Models
that are used to build the word
It is important to make a list of all parameters and record the values of these before a training or testing session .
It may be difficult to find starting values for these recogniser-specific parameters; in this case the system's manual or the company's help-desk should provide this information.