Abstrakt
This study examines ten methods of measuring short-term spectral slope with the primary aim of quantifying vowel prominence. A highly controlled dataset consisting of 45 pseudowords with varied placement of prominent vowel read by 12 speakers was used.
The investigated methods included Hammarberg index, α-measure, linear regression, differences of spectral energies in separate frequency bands, difference between amplitudes of the first two formants and spectral emphasis. All of the methods showed substantial sensitivity to speaker and vowel identity in particular.
Nevertheless, most of the methods were also able to discriminate among prominent and non-prominent realizations of the vowels. The difference between sums of energy in two adjacent frequency bands separated by a floating pivot locked to the value of the second formant proved to be the most successful measure.
On the contrary, linear regression of the spectrum performed worst. Also, exclusion of the fundamental frequency band decreased performance of the methods significantly.
The vowel /i/ was shown to be the most sensitive to prominence, whereas /u/ appeared to be the least useful for quantifying both prominence and speaker identity. A considerable speaker discrimination potential of the methods was also corroborated.