Speech-Music discriminators are usually designed under some rigid constrains. This paper deals with a more general Speech-Music Discriminator successfully used in AIDA project. The system is based on a Hidden Markov Model style classification process in which the styles are grouped into two major categories: Speech or Music. The goals of this sub-system are (1)the expandible possibilities with the addition of some new styles (like "phone female voice"), (2)the use of new rhytmical descriptors in combination with other typical ones and (3)the robustness of our speech/music discriminator in many different environments by using some Mathematical Morphology and non-linear post-processing techniques. The techniques used in our system allow a fast track in changes between styles and, thus, typical confusions in commercials can be easily cleaned. The accuracy of this system can be up to a 94.3% in broadcast radio environment.
Click to purchase paper as a non-member or login as an AES member. If your company or school subscribes to the E-Library then switch to the institutional version. If you are not an AES member and would like to subscribe to the E-Library then Join the AES!
This paper costs $33 for non-members and is free for AES members and E-Library subscribers.