Abstract:  A multiple regression model for predicting micro loudspeaker’s perceptual quality is presented in this paper.This model is based on listening test results and the objective measured data of several loudspeakers, It provides a way to know the relationship between subjective features and objective
parameters.The existing models mostly base on high-quality loudspeakers with little distortion or band limitation.Things can be very different in case of micro loudspeakers，because the distortion and band limitation can be important factors of subjective sound quality.The measuring of loudspeakers was conducted in an anechoic chamber with high resolution in frequency domain. The micro loudspeakers under test were mounted in a box which was clamped in free
space. Frequency response and total harmonic distortion (THD) were measured using sweep-frequency method. Every micro loudspeaker was measured under 0. 1 W input power for small signal condition. Several features were extracted from the anechoic measurements data. These features used in existing
models were not very suitable for describing the micro loudspeakers, because of the band limit and distortion. So the variables were modified. The listening tests were conducted in a well-controlled situation, In order to keep the experiment variables in good control，a recording was made for every loudspeaker.The recording conditions were exactly same between different loudspeakers.The recordings were made in an anechoic chamber as the same to the measurements.The listening tests were designed as a multiple comparison with graded preference response.The tests were supported by a graphical user interface program on PC. All subjects in the listening tests were well trained and with normal hearing.The listeners’task was to rate their
preference grade of each stimuli with considering spectral balance，clarity, distortion and abnormal sound，but disregarding the spaciousness.The preference ratings from each listener were analyzed using F-test. All the listeners’ data get a highly significant effect on the source of loudspeaker.
Using multiple regression analysis，a perceptual sound quality model has been constructed. Multiple regression analysis of the 6 independent variables was performed using SPSS software that calculates several possible models.The best fit model uses 3 variables producing a correlation of 0. 733，the model
accounts for 72. 5%of the variance in the observed preference rating. An ANOVA test indicated a small probability that the model product the predicted result due to chance(p<0.1).The smoothness of the frequency response curve is found contributing most to the perceptual sound quality.