Source-filter separation

Source filter separation This is one area where our model starts to break down significantly. Firstly we saw that modelling the glottis as a completely isolated source is quite umeahstie. In addition however, there are many physiological effects that make a complete source/filter separation unrealistic. For example, it is clear that when speaking with a high fundamental... 

Nearly all speech analysis is concerned with a three key problems. FirsL we wish to remove the influence of phase second, we wish to perform source/filter separation, so that we can study the spectral envelope of sounds independent of the source that they are spoken with. Finally we often wish to transform these spectral envelopes and source signals into other representations, that are coded more efficiently, have certain robustness properties, or which more clearly show the linguistic information we require. 

We will now turn to the important problem of source-filter separation. In general, we wish to do this because the two components of the speech signal have quite different and independent linguistic functions. The source controls the pitch, which is the acoustic correlate of intonation, while the filter controls the spectral envelope and formant positions, which determine which phones are being produced. There are three popular techniques for performing somce-filter separation. First we... 

Figure 12.11 shows the process of calculating the DFT, log, and inverse DFT on a single frame of speech. We will now look at how this operation performs source/filter separation. 

Linear Prediction (LP) is another technique for source filter separation, in which we use the techniques of LTI filters to perform an explicit separation of source and filter. In LP we adopt a simple system for speech production, where we have an input source x[n, which passed through a linear time invariant filter h[n], to give the output speech signal y n. In the time domain this is ... 

The preceding sections showed the basic techniques of source filter separation using first cepstral then linear prediction analysis. We now turn to the issue of using these techniques to generate a variety of representations, each of which by some means describes the spectral envelope of the speech. 

Linear prediction performs source/filter separation by assuming an HR system represents the filter. This allows the filter coefficients to be found by a process of minimising the error predicted from the HR filter. 

One of the commonalities with the formant model is that LP synthesis maintains a source/filter separation. This means that for a sequence of frames, we can res5mthesise this with a different fundamental frequency to that of the original. The benefit is that for a given transition effect that we wish to synthesise, we need only analyse one example of this we can create the full range of fundamental frequency effects by the separate control of the source. 

Perhaps the mostly widely used second generation signal processing techniques are the family called pitch synchronous overlap and add, (shortened to PSOLA and pronounced /p ax s ow 1 ax/). These techniques are used to modify the pitch and timing of speech but do so without performing any explicit source/filter separation. The basis of all the PSOLA techniques is to isolate individual pitch periods in the original speech, perform modification, and then resynthesise to create the final waveform. 

We can now ask ourselves, how does TD-PSOLA work Or in other words, after all we have said about explicit source/filter separation how is it that we have been able to change the characteristics... 

Modification is performed by separating the harmonics from the spectral envelope, but this is achieved in a way that doesn t perform explicit source/filter separation as with LP analysis. The spectral envelope can be found by a number of numerical techniques. For example, Kain [244] transforms the spectra into a power spectrum and then uses an inverse Fourier transform to find the time domain autocorrelation function. LP analysis is performed on this to give an allpole representation of the spectral envelope. This has a number of advantages over standard LP analysis in that the power spectrum can be weighted so as to emphasise the perceptually important parts of the spectrum. Other techniques use peak picking in the spectrum to determine the spectral envelope. Once the envelope has been found, the harmonics can be moved in the frequency domain and new amplitudes found from the envelope. From this, the standard synthesis algorithm can be used to generate waveforms. 

---