Study On Artificial Bandwidth Extension Of Speech Signal

In the telephone network which people use today, the transmission is generally narrowband speech with bandwidth less than4kHz. This kind of speech can ensure the intelligibility, but the quality is not as good as the corresponding wideband speech. Although there are varieties of wideband coding standards, the updates to the communication network make them cannot be achieved within a short time. People conduct the artificial speech bandwidth extension to solve this problem. This method refers to extending the missing high frequency component in the receiver and achieving wideband speech to improve the speech quality.Traditional bandwidth extension is generally based on the "source-filter" model. This kind of algorithm can achieve good results but requires a lot of prior knowledge with higher complexity. This paper proposes an audio decomposition bandwidth extension algorithm based on K-SVD. This algorithm separates and replaces the narrowband and wideband speech style characteristics to achieve bandwidth extension in another way. It can not only get a high-quality wideband speech, but also significantly reduce the use of prior knowledge.The main tasks of this thesis are as follows:(1) Overview the procedure of speech signal preprocessing, and the extraction of linear prediction coefficient, line spectrum frequency and Mel frequency cepstral coefficient.(2) Introduce traditional "source-filter" model based bandwidth extension algorithms which include wideband spectral envelope and excitation signal estimation.(3) Analysis the audio decomposition bandwidth extension algorithms which appear in recent years, including bandwidth extension based on convolution NMF and wavelet decomposition, and lead to the latter algorithm proposed in this paper.(4) Propose an artificial speech bandwidth extension method based on K-SVD, and then make some improvements on it.(5) Do objective and subjective test analysis to the proposed algorithm, including validity test, comparison test with other algorithms, and so on.