Applied Speech and Audio Processing: With matlab examples
Download 2.66 Mb. Pdf ko'rish
|
Applied Speech and Audio Processing With MATLAB Examples ( PDFDrive )
2.6. Visualisation
25 In speech analysis, as will be described in Chapter 3, many of the muscle movements which cause speech sounds are relatively slow moving, resulting in speech which slowly changes its spectral characteristics. A useful rule of thumb is that the speech signal can be assumed to be stationary, in fact pseudo-stationary, over a period of about 20–30 ms. Thus speech analysis typically requires segmentation into 20 ms long frames [3]. The stationarity requirement also extends to linear prediction (Section 5.2.1) and many other forms of analysis. When used, each must be carefully matched against the known characteristics of the audio signals which are to be handled. 2.5.2 Time-frequency resolution Moving back to the FFT, the output frequency vector, from an N -sample FFT of audio sampled at Fs Hz, contains N /2 + 1 positive frequency bins. Each bin collects the energy from a small range of frequencies in the original signal. The bin width is related to both the sampling rate and to the number of samples being analysed, Fs /N. Put another way, this bin width is equal to the reciprocal of the time span encompassed by the analysis window. It therefore makes sense that, in order to achieve a higher frequency resolution, we need to collect a longer duration of samples. However for rapidly changing signals, collecting more of them means we might end up missing some time-domain features as discussed in Section 2.5.1 and Infobox Visualisation of signals on page 32. So there is a basic uncertainty principle operating here: a single FFT can trade off between higher frequency resolution (more samples) or higher time resolution (fewer samples) but cannot do both simultaneously. Solutions vary with the requirements of the problem, but there are several frequency estimation alternatives to the FFT, and it may often be possible to perform two FFTs, over long and short analysis windows, respectively. Later in Section 6.2.2 we will describe more computationally intensive methods of attempting to satisfy both the demand of high frequency resolution and of high time resolution. Download 2.66 Mb. Do'stlaringiz bilan baham: |
Ma'lumotlar bazasi mualliflik huquqi bilan himoyalangan ©fayllar.org 2024
ma'muriyatiga murojaat qiling
ma'muriyatiga murojaat qiling