Excitation Signal Extension
This chapter deals with the estimation of the narrowband excitation signal, its extension and with finding an adequate gain factor for performing a power adjustment. Concerning bandwidth extension algorithms this represents one of two main parts as explained in Sect. 2.3. If we consider the source-filter model introduced in Sect. 2.2 the extension of the excitation signal corresponds to the process within the source path. To be more precise we are dealing with the three blocks envelope extraction, excitation signal extractionand excitation signal extensionin Fig. 2.8. Figure 4.1 shows a more detailed view of these functional blocks. In the first processing stage the narrowband excitation signal e(nb)(n) is extracted by the application of a whitening filter (predictorerror filter). This whitening filter depends on the LPC analysis of the narrowband input signal. The next functional block represents the main task within the extension of the excitation signal. Therefore we will present three major groups of methods for extending the excitation signal. They can further be distinguished in time-domain and frequency-domain approaches. The first group will be non-linear characteristics[Carl 94a, Carl 94b, Kornagel 03] which are applied in the time domain. The second one comprises so-called spectral shifting approaches(see [Kornagel 01], [Fuemmeler 01]), which can be applied in the time domain as well as in the frequency domain, followed by the third group, the so-called function generatorsthat are most often applied in the time domain. For the latter we will only present the approach where the excitation signal is modeled by white noise or sine generators [Miet 00]. Another method which is not presented in this book due to its limited results concerning speech quality is the so called fricative spreadingwhere the upper range of the spectrum is spread towards the extension region if a fricative has been detected [Heide 98],[Mason 00]. The next functional block in Fig. 4.1 performs an optional spectral whitening. This is due to the fact that some of the algorithms for extending the excitation signal presented in this book perform a spectral coloration which has to be reversed. In this chapter we also describe how to adjust the power of the estimated broadband excitation signal to the narrowband input which is represented by the last functional block in Fig. 4.1. This is a very important part since a wrong amplification or a temporal jitter in the amplification produces strong artifacts.
A detailed evaluation, including subjective as well as several objective criteria, of the different algorithms that are presented in this chapter can be found in Sect. 6.1.
KeywordsFundamental Frequency Functional Block Excitation Signal Speech Quality Spectral Envelope
Unable to display preview. Download preview PDF.