This paper describes a method for analysing the perceptual relevance of parameter variations of the Liljencrants Fant (LF) model for the glottal pulse. A perceptual distance measure based on excitation patterns was developed and evaluated in order to predict audibility discrimination thresholds for small changes to the R-parameters of the LF model. For a number of R-parameter sets, taken from real source data, an approximation Q of the distance measure was used to compute the directions of maximal and minimal perceptual sensitivity. In addition, we show that the inverse of Q can be used to calculate the amount of variation of the R-parameters to reach a just noticeable difference (JND). The results were evaluated in a listening test in which JNDs for R-parameter changes were measured. Discrimination thresholds were fairly constant across all tested conditions and corresponded on average to an excitation pattern distance of 4.3 dB. An additional error analysis demonstrated that Q is a fair approximation of the perceptual distance measure for small variations of the R-parameters up to one just noticeable difference.
|Number of pages||15|
|Publication status||Published - Jun 2004|
- previous termGlottal-pulsenext term
- Perceptualnext term distance measure
- Speech synthesis
- LF previous termparameters