Speech, silence, music and noise classification of TV broadcast material

Speech processing can be of great help for indexing and archiving TV broadcast material. Broadcasting station standards will be soon digital. There will be a huge increase in the use of speech processing techniques for maintaining the archives as well as accessing them. This paper starts with a review of several techniques used for classification of speech, music and noise. Generally, approaches that use Neural Networks (NN) or Hidden Markov Modelling (HMM) do not allow to “look inside” the network or models to determine which aspect of the sounds are similar to each other. This makes it difficult for the researcher to determine the features of the audio that are important and which ones can be ignored [1]. Furthermore, for archiving TV broadcast material, the segment time accuracy does not need to be as precise as when labelling speech corpora to be used for speech recognition research. Here, it is more important to have the correct label than to have the precise start and finish times of each segment. We present an application of information theory to the classification and automatic labelling of TV broadcast material into speech, music and noise. We use information theory to construct a decision tree from several different TV programs. This is known as the training data. We then apply this decision tree to a different set of TV programs, known as test data. We present the classification results on the training and test data sets. The correct classification rate at the frame level, for the training data was 95.5%, while for the test data it ranged from 60.4% to 84.5%, depending on the TV program type. At the segment level, the correct recognition rate and accuracy on the train data were 100% and 95.1%, respectively while for the test data the %correct ranged from 80% to 100% and %accuracy ranged from 64.7% to 100%.