Zamansal gruplandırma kullanarak sesin bant genişliğinin yapay olarak artırılması

Historically public telephone networks operate with narrowband speech, which is ban-dlimited to (250, 3400) Hz in frequency. Even though public telephone exchanges are digitaltoday, the low bandwidth limitation is still present due to the characteristics of the tradi-tional analogue network and related standards. Although intelligibility of the narrowbandspeech is high, studies show that the perceived quality of the narrowband speech is signif-icantly degraded compared to wideband speech, which is bandlimited to (50, 7000) Hz infrequency. In this thesis, we investigate the Artificial Bandwidth Extension problem, whichaims to reconstruct the missing frequency in wideband speech from narrowband speech.To solve the problem, we utilize the well-known source-filter reproduction of the humanvoice production system. This model decomposes the speech signal into two, namely thesource signal and the filter representing spectral envelope. The source signal is extendedwith up-sampling with zero insertion (spectral folding) and we propose a new framework forthe estimation of wideband spectral envelope from narrowband. The proposed frameworkbuilds temporal clusters of the joint sub-phone patterns of the narrowband and widebandspeech signals using a parallel branch HMM structure. The joint sub-phone patterns definetemporally correlated neighborhoods, in which a linear prediction filter estimates spectralfeatures of the corresponding wideband signal from the narrowband signal. The proposedframework is compared to a benchmark vector quantization based artificial bandwidth ex-tension algorithm. Objective metrics and a subjective test shows that the reconstructedwideband speech with our method significantly outperforms the narrowband speech andwideband speech reconstructed with the benchmark system.