Comparison of Time-Frequency Feature Extraction Techniques for Environmental Sound Recognition

This paper is the continuation of previously published work in which we have been analysing different methods – traditionally used in speech recognition – for their suitability to be applied to Environmental Sound Recognition. While current research devotes much effort to speech and speaker recognition, Environmental Sound Recognition is an area where little research has been reported. Despite this, environmental sound recognition is important for areas such as surveillance, because microphones need to be less focused than a video surveillance camera. This paper discusses a combinatorial experiment that investigates the use of time-frequency feature extraction techniques such as STFT and Wavelets, combined with speech recognition system learning techniques (such as the AI techniques of LVQ and ANN) for the classification of non-speech environmental sounds. This experiment reveals that a combination of a continuous wavelet transform with dynamic time warping produces the best results for environmental sound recognition. This performance is superseded only by performance on speech by Hidden Markov Models, which unfortunately are unsuitable for our purpose. Key-Words: non-speech sound recognition, environmental sound recognition, auditory signal processing, acoustic signal processing, joint time-frequency feature extraction