Laughter Detection in Noisy Settings

Spontaneous human speech contains a lot of sounds that are not proper speech, yet carry meaning, laughter being a good example. Recognizing such sounds from speech-sounds could improve speech recognition systems as well as widen the communicative range of automatic dialogue systems. Our goal is to develop methods for automatic classification non-speech vocal sounds. As laughter varies widely between individuals and cultures it represents a nice subset for studying various detection and analysis techniques for this purpose. The approach we describe here is based on the C4.5 machine learning algorithm. We focus on finding the onset and offset of laughter using single-speaker audio recordings. Prior efforts using machine learning have not, to our knowledge, used C4.5. To the best of our knowledge, our results are the best so far detecting laughter from non-laughter sounds, using a singlespeaker/single-microhpone signal with noisy background (general office environment), 89.9% at best. Here we describe our method and detail the results from two separate experiments, the first on simply detecting laughter and the second applying this method for differentiating between three different kinds of non-laughter sounds.