多媒体内容分析与检索技术概要1.ppt

* * * 控制在10页 * * * Method does not account for the frequency response of the human ear The necessary equalization can be added by applying the Fletcher-Munson equal-loudness contours Human ear can hear over 120 decibel range Software produces approximately over a 100 decibel range from a 16 bit audio recordings * 由於生理構造不同，男女生的音高範圍並不相同，一般而言： 男生的音高範圍約在 35 ~ 72 半音，對應的頻率是 62 ~ 523 Hz。 女生的音高範圍約在 45 ~ 83 半音，對應的頻率是 110 ~ 1000 Hz。 However, it should be emphasized that we are not using pitch alone to identify male or female voices. Moreover, we also use the information from timbre (or more precisely, formants) for such task. More information will be covered in later chapters. 但是我們分辨男女的聲並不是只憑音高，而還是依照音色（共振峰），詳見後續說明。 * Ex. Putting hand over mouth as you speak reduces brightness of speech as well as loudness. * Computed as magnitude-weighted average of differences between spectral components centroid * These aspects of sound vary over time. Trajectory in time is computed during the analysis but not stored in the db. Average value over trajectory Variance of value over trajectory Autocorrelation – measure of smoothness of trajectory – can distinguish between a pitch glissando a wildly varying pitch (ex) which the simple variance measure cannot 1st 3 are weighted to emphasize the important sections of the sound Table shows resulting analysis for a recording of male laughter. You can see some of the important characteristics of the sound. * * Specify sound directly by submitting constraints on the values of the N-vector directly to the system. Ex. User can ask for a certain range of pitch. OR train by example. M - # of sounds in summation In practice, can ignore off-diagonal elements of R if the feature vector elements are reasonably independent of each other. Simplifying leads to significant savings in computation time. Mean covariance together become system’s model of the perceptual property being trained by the user. * Again, off-dia