Method of modeling single data class from multi-class data

A method of modeling a single speaker from multi-speaker data, comprising the steps of:

• (a) receiving a speech signal that includes speech from multiple speakers in which speech from the single speaker accounts for more time than does speech from any other speaker;
• (b) generating a first statistical model of the received speech signal;
• (c) segmenting the received speech signal;
• (d) scoring each segment of the received speech signal using the first statistical model:
• (e) selecting a user-definable set of scores from step (d) as follows:
• a. selecting the highest 5% of the scores;
• b. discarding the lowest 5% of the scores;
• and
• c. selecting the scores between the highest 5% of the scores and the lowest 5% of the scores that are above a user-definable threshold T, where T is as follows:
• T=Min+p(Max−Min)
• where Min is the minimum score of the scores between the highest 5% of the scores and the lowest 5% of the scores, Max is the maximum score of the scores between the highest 5% of the scores and the lowest 5% of the scores, and p is a user-definable value in the range of 0≦p≦1;
• (f) identifying the segments in the received speech signal that correspond to the scores selected in step (e);
• (g) combining the segments identified in step (f);
• and
• (h) generating a second statistical model of the segments combined in step (g) as the model of the single speaker.