The fractal properties of vocal sounds and their application in the speech recognition model

In this work, we shall examine the fractal properties of simple vocal sounds such as Japanese vowels by evaluating the fractal dimension related to the self-affine property. We shall also examine the existence of chaos in the attractors reconstructed from the vocal sound waveforms by evaluating the Lyapunov exponents. The reconstructed attractors are also examined for multifractal properties. To characterize the fractal properties of complicated vocal sounds, such as speech utterances composed of several vowels, phonemes, etc., we shall propose the time-dependent fractal dimensions (TDFDs), where the fractal dimensions are evaluated based on the self-affine property, and the time-dependent multifractal dimensions (TDMFDs). We shall then use these fractal properties in a speech recognition model to examine if our method is able to characterize complicated vocal sounds effectively. For comparison, we shall utilize the running power spectrum (RPS) as a recognition parameter. It was found that utilizing the fractal properties of vocal sounds as recognition parameters gives a high recognition rate, showing that complicated vocal sounds can be effectively characterized by their fractal properties.