F0

Formants

Harmonic and Formant Amplitude

Amplitude Corrections

Energy

CPP

Harmonic to Noise Ratios

Subharmonic to Harmonic Ratio

Strength of Excitation

This section of the manual lists each of the acoustic measurements VoiceSauce is capable of making, and describes how they are made.

One of the critical measurements made by VoiceSauce is of the fundamental frequency, f0. VoiceSauce uses this measurement to estimate the location of harmonics. VoiceSauce can make measurements of F0 using four different algorithms - Straight (Kawahara et al. 1998), the Snack Sound Toolkit (Sjölander 2004), Praat, or Sun's Subharmonic-to-Harmonic Ratio method. (In VoiceSauce, these are abbreviated strF0, sF0, pF0, shrF0.) VoiceSauce defaults to the Straight measurements of F0 to locate and measure harmonic amplitudes. Straight is used to find F0 at one millisecond intervals. As of version 1.28 (December 23, 2016), the Straight algorithm in VoiceSauce is different from in all previous versions - it is now Kawahara's "XSX", taken from his new TANDEM-Straight package. The new algorithm is implemented here in a way that makes its output very similar, but not identical, to results of the old algorithm (Muliticue/"NDF"). For technical discussion and comparisons of these and other F0 estimators, see Kawahara et al. (2016); Tsanas et al. (2014).

An occasional console message, at least in older versions of VoiceSauce, is "Multicue failed: switching to exstraightsource". This somewhat cryptic message is related to the pitch estimators in the Straight package. The "Multicue" pitch estimator is quite complicated and uses cues from various calculations. However, from experience, we have found that it sometimes crashes on certain signals. When this happens VoiceSauce will switch to the simpler "extrastraightsource" algorithm which is also part of the Straight package. Since the newest Straight does not use the Multicue estimator, we expect fewer crashes.

The Snack Sound Toolkit (Sjölander 2004) is used by default to find the frequencies and bandwidths of the first four formants. By default, VoiceSauce uses the covariance method, a pre-emphasis of 0.96, and a window length of 25 ms with a frame shift of 1 ms. This frame shift is used to match the f0 estimation by the Straight algorithm.

The formant values calculated by the console version of Snack (which is
used when VoiceSauce is run under Windows) vs. the values calculated when Snack
is called via the Tcl shell (default for OSX) can differ noticeably (see https://github.com/voicesauce/opensauce-python/issues/27#issuecomment-316565993
for Terri Yu’s demonstration of some observed differences). This is because the
options implemented in the console version of Snack appear to be a subset of
those implemented in the full Snack executable, so the settings are likely to
be different. As always, inspection of obtained values for obvious errors is
recommended.

Praat can also be used to estimate formant frequencies and bandwidths. By default, Praat is configured to run with the number of formants set to 4 and the maximum formant frequency set to 6000Hz - these settings can be changed in the Settings window. As of July 2015, VoiceSauce allows Praat's fractional (x.5) values.

By default, the pitch track obtained from the Straight algorithm (Kawahara et al. 1998) is used to locate the harmonics; any of the other options may be chosen instead in the Settings window. By default, formant frequencies are obtained from the Snack Sound Toolkit (Sjölander 2004), but Praat can be specified instead in the Settings window.

In traditional FFT analysis, changing the cutting window can change the features of the extracted spectrum. VoiceSauce computes harmonic magnitudes pitch-synchronously over a three pitch period window (by default; this value can be changed under Settings). This eliminates much of the variability obtained in spectra computed over a fixed time window. The harmonics are located using standard optimization techniques which locate the maximum of the spectrum around peak locations, as estimated by F0. This is equivalent to using a very long FFT window.

VoiceSauce measures the amplitudes of various harmonics: H1, H2, H4, A1 (the harmonic nearest F1), A2, A3, H2K (the harmonic nearest 2000 Hz), H5K, H1-H2, H2-H4, H1-A1, H1-A2, H1-A3, H4-H2K and H2K-H5K, as well as corrected versions of all these measures (except for H5k). The individual harmonic amplitudes are not normalized, and will vary with, e.g., overall loudness. For this reason, the harmonic difference measures like H1-H2 are more commonly used, as they provide a kind of within-token normalization.

Note that the reliability of these measures depends on the successful estimation of their component parameters. If the F0 is not well-tracked, then all the measures that include H1 (or any other harmonic) will be problematic. Similarly, if one or more formants are not well-tracked, then the corresponding measures will be problematic. Thus if the estimate of F1 is wrong, then A1 and H1-A1 will be wrong too, even for the uncorrected measures. Errors in F1 estimation are especially likely for breathy, nasal, or high-pitched vowels. Obviously, all the amplitude corrections described in the next section also crucially depend on accurate formant estimation. Therefore it is recommended that the F0 and formant estimates be checked to verify the integrity of the voice measures derived from them.

Note that there are no formant-corrected version of the 5K measure. This was due to the observation that formant estimation inaccuracies increase significantly with the higher formants. Also, the sample rate of the input file needs to be higher than 10KHz in order to return a meaningful 5K measure.

All the harmonic-amplitude voice measurements can be corrected for the effect of formant frequencies, using an algorithm developed by Iseli & Alwan (2004, 2006, 2007). This is done so that voice parameters can be compared across segments with different formant frequencies, e.g. different vowel qualities. A formant boosts the amplitude of any nearby harmonic(s), so uncorrected values of harmonic amplitudes reflect both the source and the filter. Uncorrected outputs should be used only when comparing matched speech samples for which the filter functions will be essentially the same. In some studies, only segments with a very high F1 (e.g. low vowels) and relatively low F0 are used, so that the H1 and H2 frequencies will be well below the F1 frequency and uncorrected H1-H2 should be unaffected by the formants. Under this method, uncorrected parameters based on harmonics above H2 cannot be used, and most vowel qualities cannot be studied. The alternative of applying the formant corrections allows any mix of segments (or at least, any mix of voiced oral sonorants), with any formant frequencies, to be combined together for comparisons across the full range of harmonic amplitude measures. Formant-corrected harmonic-amplitude measures are thus especially important when using natural speech samples in which segment sets cannot be controlled.

Amplitude measures are corrected every frame using the formant frequencies obtained by default from the Snack toolkit, or from Praat if specified by the user (under Settings). Formant bandwidths are by default calculated by the formula from Hawks & Miller (1995). That is, the formant bandwidths estimated by Snack and Praat and included in VoiceSauce's output are not used by default in the corrections. Under Settings, the "Bandwidth" setting gives the option of switching to the estimated values of the formant bandwidths - from either Snack or Praat, whichever is used for the formant frequencies. Here is how the various amplitude measures are corrected:

H1* - uses F1, F2
and B1, B2 (from formula, or calculated)

H2* - uses F1, F2
and B1, B2 (from formula, or calculated)

H4* - uses F1, F2
and B1, B2 (from formula, or calculated)

H2k* - uses F1,
F2, F3 and B1, B2, B3 (from formula, or calculated)

H5k - only
uncorrected is available

A1* - uses F1, F2
and B1, B2 (from formula, or calculated)

A2* - uses F1, F2
and B1, B2 (from formula, or calculated)

A3* - uses F1,
F2, F3 and B1, B2, B3 (from formula, or calculated)

The measures are then smoothed with a moving average filter with a default length of 20 milliseconds. As noted in the previous section, if the estimates of the formant frequencies are not accurate, these corrections will be problematic.

It has been noted that formant bandwidths calculated by formula, while overall more reliable than using token-specific measurements, can cause large errors in corrections of H4 and H2k when these happen to be at a formant frequency: the harmonic amplitudes are reduced to too-low values. Using measured bandwidths, if those are reasonable, eliminates this problem. VoiceSauce was modified in June 2015 to allow such use of measured bandwidths, as described above.

In the literature, corrected measures are indicated with an asterisk. For example, H1*-H2* is the standard way of indicating the corrected form of H1-H2. However, because of limitations imposed by Matlab, asterisks cannot be used in VoiceSauce's output. Instead, a "c" indicates a corrected measure. Thus, "H1c" is H1*. For clarity, a "u" is used for uncorrected measures. Thus, "H1u" is uncorrected H1.

Energy refers to the Root Mean Square (RMS) energy, calculated at every frame over a variable window equal to five pitch pulses by default. The variable window effectively normalizes the energy measure with F0 to reduce the correlation between them.

Ceptral peak prominence (CPP) calculations are based on the algorithm described in Hillenbrand et al. (1994). A variable window length equal to 5 pitch periods is used by default for the calculations. After multiplying the data with a Hamming window, the data is then transformed into the real cepstral domain. The CPP is found by performing a maximum search around the quefrency of the pitch period. This peak is normalized to the linear regression line which is calculated between 1 ms and the maximum quefrency.

Harmonic-to-noise ratio (HNR) measures are derived from the algorithm in de Krom (1993). Using a variable window length equal to 5 pitch periods by default, the HNR measurements are found by liftering the pitch component of the cepstrum and comparing the energy of the harmonics with the noise floor. HNR05 measures the HNR between 0-500Hz, HNR15 measures the HNR between 0-1500Hz and HNR25 measures the HNR between 0-2500Hz.