Oscillations, Waves, and Interactions - GWDG

3.4.2 Analysis of running speech
Speech research 31
The analysis was only carried out on connected intervals classified as voiced of minimum
length 70 ms. A weighting by length is implemented, since longer intervals are
more expressive. In these intervals period markers are set (Fig. 3 (11)) and acoustic
quantities are determined, for instance:
• period lengths by the waveform matching algorithm;
• jitter (3 definitions), shimmer (3 definitions), MWC, GNE.
From these again a GHD can be constructed. The position of voices in the GHD is
different in running speech from that for stationary vowels, so that a new calibration
is required to obtain comparable representations for both cases. The definition of the
axes, which is based on a principal-component analysis in a high-dimensional space,
has to be carried out anew. Here, the choice of the underlying quantities was the same
for consistency reasons, but their weighting was different. The new GHD is called
“GHDT”, “T” meaning “text”. The coordinates in the GHDT are averaged over the
analyzed intervals of the text utterance, weighted by their lengths. The variances
of the measurement points in the GHD are, because of sound dependence, of course
larger than for stationary vowels, but the mean values retain their expressiveness.
The consistency of the GHDT was checked with various normal and pathological
voices and different utterances.
Besides the GHD, the automatic voiced/unvoiced classification can also be applied
to other diagnostically useful quantities in order to extend their usage to running
speech. This concerns, for instance, the Pitch Amplitude (PA; 1 st maximum of the
autocorrelation function of the prediction error signal) and the Spectral Flatness
Ratio (SFR; logarithm of the ratio of geometric and arithmetic means of the spectral
energy density of the prediction error signal).
Based on the acoustic quantities, group analyses of various phonation mechanisms
and cancer groups (significant group separation) can be conducted. For preliminary
and recent presentations of methods and results see Refs. [25–28].
So far, no phonemes were to be recognized but only their linguistic (not actual)
voicedness. Meanwhile, the perceptron method has been extended to recognition of
the six stationary vowels, using 6 output cells. Training was done with 8192 vowels
of at least 2 s duration from all kinds of voice quality. This can help to further
automatize the determination of voice quality.
3.5 Analysis of glottal oscillation
The voice pathologies are related to the functioning of the vocal folds, which form a
self-oscillating nonlinear mechanic and aerodynamic system driven by the glottal air
flow. In order to relate the acoustic voice characteristics to properties of the glottal
oscillation, these must be (if possible, automatically) recorded and characterized by
few quantities. Here, acoustic as well as optical methods are employed. These methods
have not yet been extended to running speech, but the only essential difficulty
to do so appears to be the large amount of data occurring then.