F. Salehian Matikolaie, C. Tadj
This study proposes using a novel combination of short-term and long-term features from different timescales to develop an automatic
newborn cry diagnostic system to differentiate the cry audio signals
(CASs) of healthy infants from those with respiratory distress syndrome
(RDS). Mel-frequency cepstral coefficients (MFCCs) were used as the
short-term features, while the melody and rhythm features obtained from
longer timescales were used as the long-term features. We hypothesized
that the differences between these groups may occur on several
timescales. Finally, a support vector machine model was used to
generate the final classification. Among other findings, the best
results were obtained from the combination of all three feature sets
(the MFCCs and the rhythm and melody features) in the expiration
episode; the combination of MFCCs and tilt features improved the
classifier performance in the inspiration episode. In terms of F-score
measure, in the inspiration experiment, the tilt features alone were
the strongest classification features for differentiating infants with
RDS from healthy infants. The results indicate that the combination of
short-term and long-term features provide a better classification
method for differentiating the CASs of healthy infants versus RDS
infants. Moreover, the results confirmed the importance of long-term
features in expiration and inspiration episodes as diagnostic markers
between groups of healthy infants and RDS infants.
Keywords: Long-term
features, Melody, Rhythm, Short-term features, Mel-frequency cepstral
coefficient, Support vector machine, Newborn infant cry, Expiration and
inspiration cry.