Separation of Simultaneous Word Sequences Using Markov Model Techniques
Abstract
This thesis develops a method of separating multiple simultaneous conversations through the use of Markov models. Test samples which represent the conversations to be used as training data are described by a grammar base upon word and word-pair occurrences within the text. This grammar is then used to establish a Markov model for the text. These models are then combined to form a Markov model which describes the simultaneous occurrence of multiple conversations. Artificially generated word sequences which have the same grammar as the training conversations are supplied as input to the conversation filter, whose purpose is to listen to one of the input sequences. The conversation filter takes on either an optimal form in which the grammars of all input sequences to the filter are known, or a sub-optimal form which uses only the grammar of the desired output sequence. The conversation filter utilizes the Viterbi algorithm to extract the optimal text sequence for a best match to the grammar of the desired output. Analysis is performed to determine the efficiency of the algorithm and the performance of the algorithm for varying degrees of similarity between the grammars being separated.
Document Details
- Document Type
- Technical Report
- Publication Date
- Sep 01, 1990
- Accession Number
- ADA243374
Entities
People
- James L. Kingston
Organizations
- Naval Postgraduate School