Signal analysis of the female singing voice: Features for perceptual singer identity

Individual singing voices tend to be easy for a listener to identify, particularly when compared to the difficulty of identifying the performer of any other musical instrument. What cues does a listener use to identify a particular singing voice? This work seeks to identify a set of features with which one can synthesize notes with the vocal quality of a particular singer. Such analysis and synthesis influences computer music (in the creation of synthetic sounds with different timbre), vocal pedagogy (as a training tool to help singers understand properties of their own voice as well as different professional-quality voices), and vocal health (to identify improper behavior in vocal production).; The problem of singer identification is approached in three phases: signal analysis, the development of low-order representations, and perceptual evaluation. To perform the signal analysis, a high-resolution time-frequency distribution is applied to vowel tokens from sopranos and mezzo-sopranos. From these results, low-order representations are created for each singer's notes, which are used to synthesize sounds with the timbral quality of that singer. Finally, these synthesized sounds, along with original recordings, are evaluated by trained listeners in a variety of perceptual experiments to determine the extent to which the vocal quality of the desired singer is captured.; Results from the signal analysis show that amplitude and frequency estimates extracted from the time-frequency signal analysis can be used to re-create each signal with little degradation in quality and no loss of perceptual identity. Low-order representations derived from the signal analysis are used in clustering and classification, which successfully clusters signals with corresponding singer identity. Finally, perceptual results indicate that trained listeners are, surprisingly, only modestly successful at correctly identifying the singer of a recording, and find the task to be particularly difficult for certain voices and extremely easy for others. Listeners also indicate that the majority of sounds synthesized with the low-order representations sufficiently capture the desired vocal timbre. Again, the task is easy for certain voices and much more difficult when evaluating other singers, consistent with the results from the original recordings.