Sound Pattern Classification

PI: Yang Cai
Member: Rafael de M. Franco
Contact : ycai@cmu.edu

Goals

Sound classification is an affordable sensing technology for monitoring physical intrusions. It can be applied to many situations such as board patrol. The goal is to classify sound patterns from natural environments, such as cars, trucks, airplanes, etc.

Approach

The input is a sound sample recorded by a microphone. The sample is pre-processed and the spectrogram is calculated with the FFT transform.

Bayes decision theory is applied in order to classify the signal.

Results

Three classes of sounds were used to train the program: car, airplane and helicopter. In the figure, the blue points in form of diamonds are the tested sounds for each 3 classes. The highest probability among these 3 results is the optimal detection.

Applications

• Intelligent edge protection systems
• Smart homes for elderly
• Medical diagnostics
• Ambient Intelligence systems
• Multimodal human-computer interfaces
• Safety alert systems
• Intelligent multimedia processing
• Smart phones
• Emotional computing
• Autonomous discovery systems