Towards an understanding of speech and song perception

The human singing voice plays an important role in music of all societies. It is an extremely flexible instrument and is capable of producing a tremendous range of sounds. As such, the human voice can be hard to classify and poses a major challenge for automatic audio discrimination and classification systems. Speech/song discrimination is an implicit goal of speech/music discrimination, where a division is sought between speech and song, such that the singing voice can be grouped together with other musical instruments in the same category. However, the division between speech and song is unclear and even human attempts at speech/song discrimination can be highly subjective and open to discussion. In this paper we present the results of a test that was designed to investigate differences in auditory perception for speech and song. Twenty-four subjects were instructed to attend to either the words or pitch, or both words and pitch of context-free spoken and sung phrases. After presentation of each phrase, subjects were asked to either type the words that they recalled, or select the correct pitch contour from a choice of four graphical representations, or do both, depending on the task specified before presentation of the phrase. The results of the experiment show a decrease in the amount of linguistic information retained by subjects for sung phrases and also a decrease in accuracy of response for the sung phrases when subjects attended to both words and pitch instead of words or pitch alone.