Oscillations, Waves, and Interactions - GWDG

3.4 Transfer to running speech
Speech research 29
The voice analysis was originally based on stationary vowels. In clinical diagnostics,
however, a voice analysis from continuous speech is required in order to objectively
assess the vocal disease under normal stress and to be able to treat it optimally.
The stationary phonation corresponds rather to a singing voice, contrary to the
more natural running speech. Thus for a comprehensive description of voice quality
the analysis of running speech is an essential extension of the analysis of stationary
phonation. The methods of vowel analysis should be partially transferable to voiced
intervals in running speech. For this purpose a method was developed to recognize
such intervals automatically. The main difficulty was that the linguistically voiced
sounds are not necessarily realized as voiced for strong voice disorders.
3.4.1 Determination of voiced and unvoiced intervals
A voiced/unvoiced classification by (e. g.) zero-crossing and correlation techniques
directly on the speech signal would, for strongly disturbed voices, recognize too few
voiced intervals. Thus a consideration of the spectral envelope (formant structure) is
preferable, which little depends on the actual glottal excitation. The method uses a
3-layer perceptron with sigmoid activation function (values 0 to 1) as classifier. The
template vectors for its input were formed as follows (numbers refer to Fig. 3):
The speech signals, digitized with 48 kHz, were downsampled to 12 kHz and decomposed
into overlapping Hann-windowed 40 ms intervals with 10 ms frame shift (3,4).
Pauses are eliminated based on an empirical energy threshold. An LPC analysis of
12 th order (autocorrelation method; preemphasis 0.9735) yields a model spectrum
(5), which is converted to 19 critical bands (Bark scale) by summation in overlapping
trapezoidal windows (6). It is compressed with exponent 0.23 and normalized
by its maximum over time and critical bands (7,8). The LPC order and method were
optimized to yield minimal misclassification.
The optimal values of the perceptron parameters (number of hidden cells, learning
rate, classification threshold, number of iterations, training material) were determined
in extensive experiments (6750 different cases, about 12000 spectra). Twelve
hidden cells worked best. As training method of the perceptron (9), an accelerated
backpropagation [24] was employed with learning rate 0.01 and momentum term 0.8.
The classification threshold at the output is 0.45, the desired net outputs for training
are 0.1 (unvoiced) and 0.9 (voiced). The weights are initialized with random numbers
in the range 0 to 1. Three perceptrons with different initial weights were used
in parallel, averaging their recognition scores.
Since our own speech data were unlabeled, they could not serve for training. Instead,
the training set consisted first of 32 phonetically segmented texts (16 times
“Nordwind und Sonne”, 16 times “Berlingeschichte”) from 16 different normal speakers
in the German Phondat database, a total of 154550 labeled Bark spectra, excluding
pauses. The training used up to 5000 iterations. For testing, different subsets of
30 of the 32 texts were used in training and the remaining 2 in the test. The error
score amounted to 4.8%. With the above threshold (0.45), only 25% of these are
falsely classified as voiced; false unvoiced classification is less detrimental. As mis-