Abstract
Abstract Sounds play an important role in our life. We can communicate with each other and know what happens by listening to sounds. By extracting the feature of sounds, we can keep specific information of sounds to recognize sounds. Sound transmission can be done if sounds could be reconstructed from the transmitted features of sounds. In this research, we attempt to reconstruct sounds using features that are typically transmitted for recognition purposes. In this thesis, we take the mel frequency cepstral coefficients (MFCC), a set of features that has been commonly used for sound recognition, as the basic features for reconstruction. Because MFCC does not encode the detail of sounds, we use the pitch as additional information to enhance the completeness of the features. The sound reconstruction is based on a source-filter model which takes the reconstructed frequency response from MFCC as the spectral envelope and determines the sound source with the pitch. The critical factors of the reconstructed sound source are the frequency distribution of noise and harmonics which could be determined by the human speech production mechanism. We then combine the spectral envelope with the sound source to reconstruct sounds through a modified source-filter model. In this thesis, we test our methods by analysis and reconstruction of speech and non-speech materials. We attempt to find the factors that may affect the quality of reconstructed sounds. We also evaluate reconstructed sounds by subjective listening test and objective perceptual evaluation of audio quality ( PEAQ). The range of grades is from 1(very bad) to 5(very good). The result of listening test reveals that the grade of speech and non-speech reconstruction is about 3 to 4. PEAQ reveals that the grade of non-speech reconstruction is about 2 to 3.5 and the grade of speech reconstruction is slightly higher than 1.