Salient Features for Anger Recognition in German and English IVR Portals

Abstract

Anger recognition in speech dialogue systems can help to enhance human computer interaction. In this paper we report on the setup and performance optimization techniques for successful anger classification using acoustic cues. We evaluate the performance of a broad variety of features on both a German and an American English voice portal database which contain "real" speech, i.e. non-acted, continuous speech of narrow-band quality. Starting with a large-scale feature extraction, we determine optimal sets of feature combinations for each language, by applying an Information-Gain based ranking scheme. Analyzing the ranking we notice that a large proportion of the most promising features for both databases are derived from MFCC and loudness. In contrast to this similarity also pitch features proved importance for the English database. We further calculate classification scores for our setups using discriminative training and Support-Vector Machine classification. The developed systems show that anger recognition in both English and Salient Features for Anger Recognition in German and English IVR PortalsGerman language can be processed very similarly reaching comparable results.

Anger recognition in speech using acoustic and linguistic cues

“…On the basis of a voiced, unvoiced and silence segmentation we calculate ratios of features from these segments both separately and jointly. A more detailed description of the feature setup can be found in Polzehl et al (2010).…”

Automatically Assessing Personality from Speech

“…We extract audio descriptors using a 10ms frame shift, and derive statistics from these descriptors at the utterance level. Overall, we generate about 1450 features, which we have successfully used in our previous work on emotion recognition [6]. Features are extracted using Praat 1 .…”

“…Best combinations of features were discovered afterwards, when more features were taken into account and subsets were compiled by diverse methods [6].…”

Automatically assessing acoustic manifestations of personality in speech

“…For our classifier, we will rely on prosodic and acoustic features, in line with findings of salient features reported in related work [6,7]. We leverage from previous work on emotion recognition [12], and extract audio descriptors such as 16 MFCC coefficients, 5 formant frequencies, intensity, pitch, perceptual loudness [13], zero-crossingrate, harmonics-to-noise-ratio, center of spectral mass gravity (centroid), the 95 % roll-off point of spectral energy and the spectral flux, etc, using a 10 ms frame shift. From these descriptors, we derive statistics at the utterance level, separate for voiced and unvoiced regions, on speech parts only.…”

“…A more detailed description of the total of 1450 features is beyond the scope of this paper, and can be found in [12].…”