ASR performance.

Some of the main external factors influencing the performance of the vocabulary dependent recognizers are gathered here to categorize the applications and understand the limits of the state-of-art ASR products.

1.

The size and complexity of the vocabulary naturally affect the difficulty of the task.

2.

Task based constraints that reduce the set of word candidates for each utterance are commonly used extensively to provide opportunities for large vocabulary speech recognizers. The average word branching factor is called perplexity that gives an idea of the expected number of words between which the recognizer must actually make its decision.

3.

The data collecting conditions may require special signal processing tricks, if there is much noise in the environment or if the conditions are instable.

4.

The speaking rate and the silence between successive words is one fundamental criterion to classify the ASR tasks. Speech used in natural conversations between humans is very difficult to recognize.

5.

The match between the training and testing material is important. The more difference there is between the data, the more re-training or adaptation is required for good performance.

6.

Speaker dependence divides the tasks into different categories. At one end good performance for one speaker can be obtained with a relatively few training data and simple models. At the other end the training of models which should perform adequately for anyone needs serious efforts.