A Pitch Extraction System Based on Laryngeal Mechanisms Classification

摘要

Pitch extraction is one of the most important areas in signal processing. One of the reasons for this fact is because it is a key component of several speech processing, coding or synthesis systems. Many methods were proposed to date; however, there are possible improvements yet, mainly concerning the fine-tuning of parameters, since the majority of the works focus on the definition of completely new approaches. This paper proposes a system to estimate pitch based on laryngeal mechanisms classification. Currently, this classification is based on texture discrimination between visual representations of the audio signal. Therefore, at first, we improve state-of-the-art accuracy in classifying laryngeal mechanisms through changing this image generation process, based on the spectrogram of the signal, and by tuning the used classifier. After that, we optimize the frequency range of the pitch detectors relying on that classification. Such optimization is possible because each laryngeal mechanism has a frequency range, and therefore, it is not necessary to use the complete range of human voice to every sound; it was produced in just one specific mechanism. Since our system consists of parameter tuning, it is not limited to any specific pitch extraction method. Our experiments over two well-known pitch detectors show that this adjustment on the frequency range based on laryngeal mechanism can significantly improve pitch detection accuracy.