Speech Recognition Silence Cluster Seeding
Original Publication Date: 1984-May-01
Included in the Prior Art Database: 2005-Feb-04
Clusters of silence in acoustic prototypes used for continuous speech recognition are automatically controlled by an algorithm which performs a pre-scan to make a histogram of energies of observation, fits the histogram with two normal distributions by iterations of Markov estimation procedures to yield means, variances and proportions, and selects cluster seeds to decrease the amount of either speech or silence. For example, 90% speech/10% silence might be chosen. Clustering algorithms used in the speech recognition process use randomly selected seeds; non-speech (silence) is thus represented in the seeds proportionally to the amount of silence in that particular training utterance. This invention measures the amount of silence, and adjusts the seeds so that the silence is represented at a prespecified level.