Measures of spectral tilt

Abstract

Crude measures of spectral tilt (F0-H2 difference, and F0-F1 difference) have been demonstrated to be useful for distinguishing phonation types. However with such methods, it is difficult to control for differences due to variations in vowel quality and F0. In order to place such measures on a firmer foundation, the differences in vowel quality can be compensated for by inverse filtering. This technique has been used for analyzing vowels in languages having contrasting phonation types. FM recordings of airflow data in Burmese and Hmong, and ordinary AM audio recordings of !Xóõ and Jalapa Mazatec were analyzed. AM recordings can be used, as phase distortion may be neglected while working in the frequency domain. FFT spectra were made of the inverse-filtered waveforms. We considered several questions such as computational methods for deriving the amplitudes of harmonics, the expected dependency of amplitude on frequency, and the appropriate range of frequencies to examine. The results show that measures of spectral tilt obtained from inverse-filtered data can be used to characterize differences in phonation type. [Work supported by NINCDS.]

Measures of phonation type in Hmong

“…These results remind us that care should be taken in making comparisons between results of modeling studies and_measures of real speech. Measures of spectral tilt are considered further inJackson et al (1985).Figure 10 shows mean values and standard deviations averaged over speakers for another measure of spectral balance, the difference in amplitude between the first and second harmonics (F 0 -H 2 ). The nonbreathy tones have very similar values for this difference, all clustering around slightly over 2 dB.…”

FD‐based detector for medical image watermarking

“…The test statistic for the PD can be given below: Remark 1 Selection of fractional order q The transfer function of the FD is given by (18). The magnitude of the transfer function is given as follows: Differentiating (18), we get the following expression: Equation (19) is called spectral tilt of the filter [33]. Most often in signal and speech processing application, high frequency contains low‐power spectral density.…”

“…A total number of ‘seeds’ are 1024. 5: For every block, fix the mid‐frequency range index, where the watermark signal has to be added. In our case, we consider the indices [5–7, 12–14, 19–21, 26–28, 33–35]. 6: The size of DSSS is 15.…”

Comparison of spectral tilt measures for sentence prominence in speech—Effects of dimensionality and adverse noise conditions

“…Other studies use an array of different methods, including calculation of the difference in dB between the overall intensity and the intensity of the fundamental frequency (or of the intensity in a frequency band centered at F0) (Barbosa, Eriksson, & Åkesson, 2013;Eriksson, Thunberg, & Traunmüller, 2001;Heldner, 2001), taking the first cepstral coefficient (C1) (Tsiakoulis, Potamianos, & Dimitriadis, 2010), taking the difference in dB between a signal with high-frequency pre-emphasis and flat frequency weighting (SPLH-SPL) (Fant, Kruckenberg, Liljencrants, & Hertegård, 2000), taking the difference in dB between the first harmonic and third formant (H1-F3) (Okobi, 2006), fitting a regression line in the magnitude spectrum (Aronov & Schweitzer, 2016;Lu & Cooke, 2009), taking the band-limited spectral energy ratios (Murphy, McGuigan, Walsh, & Colreavy, 2008;Prieto & Ortega-Llebaria, 2006), using the long-term average spectrum (LTAS) to obtain band-limited energy ratios (Sundberg & Nordenberg, 2006), and using all-pole modeling techniques (Magi, Pohjalainen, Bäckström, & Alku, 2009). In addition, some studies utilize similar measures, such as regression line fitting and harmonic ratio, but, instead of applying the measures directly on the short-term spectrum of speech (such as in the case of SUT), they utilize the spectrum of the glottal source waveform obtained through glottal inverse filtering (GIF) (see, e.g., Iseli et al, 2006;Jackson, Ladefoged, Huffman, & Antoñanzas-Barroso, 1985;Kreiman, Gerratt, & Antoñanzas-Barroso, 2007). Other studies make use of various parameterizations of the voice source, such as the Liljencrants-Fant (LF) model (Fant, Liljencrants, & Lin, 1985), in order to derive a measure for tilt (see, e.g., Fant & Kruckenberg, 1994) and may also use other parameters of the voice source in order to study and evaluate different prosodic phenomena (see, e.g., Fant & Kruckenberg, 1994;Iseli et al, 2006).…”