Applied Speech and Audio Processing: With matlab examples
Download 2.66 Mb. Pdf ko'rish
|
Applied Speech and Audio Processing With MATLAB Examples ( PDFDrive )
Speech communications
5.2.1.3 Pre-emphasis of the speech signal An important practical point to note when performing LPC analysis is that the LPC coefficients that are found are supposed to match the analysed speech signal as closely as possible. However it turns out that the LPC equations tend to satisfy the lower fre- quencies while matching the higher frequencies more poorly. Thus it is common to emphasise the higher frequencies prior to LPC analysis. In fact, when speech is radiated from a human mouth, from an area of high pressure, through a constriction, into a low pressure area, a spectral roll-off occurs to reduce the amplitude of the higher frequencies. Thus speech recorded from outside the mouth will differ from speech recorded from inside the mouth (and there do exist tiny microphones that can record speech from within the mouth). Since the LPC filter employed in speech analysis is supposed to model the vocal tract response, it is preferable to allow it to analyse the signal produced by the vocal tract without the influence of lip radiation. We can therefore counteract the effects of lip radiation by performing pre-emphasis before LPC analysis. Given a speech signal s (n), the pre-emphasis would normally be performed with a single-tap filter having transfer function (1 − αz −1 ), and an emphasis coefficient, α of nearly 1. A typical value used in research systems is α = 15/16 = 0.9375. Thus each pre-emphasised speech sample s (n) comes from the current and previous input speech samples acted on by the following FIR filter: s (n) = s(n) − 0.9375 × s(n − 1). (5.3) By performing pre-emphasis of the speech signal in this way prior to LPC analysis, we can better approach the signal that leaves the vocal tract, and can overcome one of the issues with LPC analysis where the coefficients match the higher frequency components poorly. Of course, any speech system that outputs something based upon pre-emphasised speech will sound a little strange – we do not normally hear speech from within a person’s mouth. So even though the processing can be conducted on pre-emphasised speech, the output must be de-emphasised, to replace the attenuated low frequencies which we had removed. The de-emphasis filter matches the emphasis filter, reversing the emphasis applied to the speech signal s (n) to recreate more natural sounding speech r(n). This IIR filter is as follows: r (n) = s (n) + 0.9375 × r(n − 1). (5.4) In Matlab we can create a pre-emphasis filter very easily, and apply this using the filter() function. If the original speech signal is called s, the pre-emphasised output is to be es and the de-emphasised version of this is ds then we can easily convert between them: % Create the emphasis/de-emphasis filter coefficients |
Ma'lumotlar bazasi mualliflik huquqi bilan himoyalangan ©fayllar.org 2024
ma'muriyatiga murojaat qiling
ma'muriyatiga murojaat qiling