The effect of low-pass filtering on estimated voice source parameters

effect of low-pass filtering exactly is. This was the first Voice source parameters are often obtained by are not optimally suited for this task (see e.g. [6]). Finding parametrizing glottal flow signals. However, before an appropriate evaluation method was the second goal of parametrization these glottal flow signals are usually lowour research. The evaluation method should make it pass filtered. As low-pass filtering changes the shape of possible to determine the magnitude of the errors in the the glottal pulses, it will also cause an error in the estimated voice source parameters that are due to lowestimated voice source parameters. The present article pass filtering. Finally, the third goal of our study was to presents results of our research on the effect of low-pass develop an estimation method for which these errors are filtering on the estimated voice source parameters. small. To this end three parametrization methods are We will first present an evaluation method which makes it possible to study the effect of low-pass filtering in detail. 1. METHOD The evaluation results show that low-pass filtering leads to an error in all estimated voice source parameters. Parametrization of dU or U can be done in several ways. However, the magnitude of the errors differs for the Usually landmarks (like minima, maxima, zero crossings) various voice source parameters, and also depends on the are detected in the signals. Because these landmarks are estimation method used. We will show that the errors can estimated directly from the voice source signals, these be reduced substantially by choosing the appropriate methods will be called direct estimation methods (DE estimation method. methods). INTRODUCTION voice source model to the data. Because in estimation The technique of inverse filtering has been available for a they will be referred to as 'fit estimation' methods (FE long time now. Inverse filtering can be used to obtain an methods). As a voice source model we use the LF model estimate of the first derivative of glottal flow (dU ). [7]. In our FE method five LF parameters (E , t , t , t , and g Subsequently, the effect of lip radiation can be canceled T ) are estimated for each pitch period. The FE method by integrating dU to obtain an estimate of true glottal consists of three stages: g flow (U ). However, estimating a voice source signal 1. initial estimate g (either dU or U ) is usually not enough. For many 2. simplex search algorithm g g applications it is necessary to parametrize the glottal flow 3. Levenberg-Marquardt algorithm signals. Before the glottal flow signals are parametrized, they are in [4], primarily because the authors provide a fairly low-pass filtered at least once in all methods, viz. before detailed description of their method (see especially page A/D conversion. Often, they are low-pass filtered again 765 of their article), and because with this method it is after A/D conversion, usually to cancel the effects of possible to estimate the LF parameters E , t , t , and t . In formants that were not inverse filtered or to attenuate the their method Alku & Vilkman [4] do not estimate T . noise component. However, low-pass filtering changes the Since an LF model is not complete without T , T was shape of the glottal flow signals, and, consequently, estimated by fitting the second part of the LF model to the influences the estimated voice source parameters [1, 2, 3, return phase of dU . Therefore, strictly speaking, only E , 4, 5, 6]. t , t , and t can be said to be the result of the DE method, Therefore, it becomes important to determine what the procedure. goal of the present study. Previously proposed methods