Comparing measurement errors for formants in synthetic and natural vowels.

摘要: The measurement of formant frequencies vowels is among the most common measurements in speech studies, but are known to be biased by particular fundamental frequency (F0) exciting formants. Approaches reducing errors were assessed two experiments. In first, synthetic constructed with five different first (F1) values and nine F0 values; bandwidths, higher frequencies, constant. Input compared manual automatic measures using linear prediction coding-Burg algorithm, closed-phase covariance, weighted prediction-attenuated main excitation (WLP-AME) algorithm [Alku, Pohjalainen, Vainio, Laukkanen, Story (2013). J. Acoust. Soc. Am. 134(2), 1295-1313], spectra smoothed cepstrally averaging repeated discrete Fourier transforms. Formants also measured manually from pruned reassigned spectrograms (RSs) [Fulop (2011). Speech Spectrum Analysis (Springer, Berlin)]. All WLP-AME RS had large direction strongest harmonic; smallest occur RS. second experiment, these methods used on isolated words spoken four speakers. Results for natural show that bias affects all methods, including WLP-AME; only formants appeared accurate. addition, coped better weaker glottal fry.