Paper Title
Detection Of Ambulance And Fire Truck Siren Sounds Using Neural Networks

Abstract
Automatic methods for detecting ambulance and fire truck siren sounds are investigated in this work. In general, emergency service vehicle such as ambulances, fire trucks, and police cars use sirens to warn other road users for quick passage, especially for moving through the traffic. However, due to the well soundproofing techniques in modern cars, drivers may not be aware of the approach of emergency vehicles, especially when in-vehicle audio systems are used. As a consequence, emergency vehicles may be blocked and even collided with other vehicles. To help drivers avoid being unaware of approaching emergency vehicles, this work proposes automatic detection methods of siren sounds. But during the initial study stage, we focus only on the siren sounds of ambulance and fire truck in Taiwan. The detection task is formulated as a problem of identifying the three sound classes, respectively from ambulance, fire truck, and ambient noise. We propose to use recurrent neural networks to determine which of the classes each 1-second sound recording belongs to, based on its Mel-frequency cepstral coefficients. Our experiments shows that the proposed method can achieve the identification accuracy of 90% in simulated -15dB noisy sound data and 93.8% in real sound data recorded on a downtown street. Index terms - Siren sound detection, neural network, Multi-layer perceptron, Long-Short Term Memory RNN, Mel-Frequency Cepstral Coefficients.