Investigations of the potential of acoustic speech regions for detecting hostile talking in Algerian Dialect and German language

• Published in: Applied acoustics 2022-06, Vol.195, p.108820, Article 108820
• Main Authors: Ykhlef, Faycal, Benzaba, Wahiba, Boutaleb, Ratiba, Bouchaffra, Djamel, Derbal, Abdeladhim
• Format: Article
• Language: English
• Subjects: Acoustic Attributes; Algerian Dialect; Emotion Recognition; German Language; Hostile Talking Detection; Speech Regions
• ISSN: 0003-682X; 1872-910X
• DOI: 10.1016/j.apacoust.2022.108820

•A methodology for detecting anger based on speech waves was introduced in Algerian Dialect (AD) and German Language (GL)•The efficiency of speech regions to spot anger differs from one language to another.•A speech emotional corpus of AD was recorded.•The detection of anger in AD is more effective when voiced waves are used.•Transition regions show a slightly better performance for detecting anger in GL compared to voiced waves. To find out what are the acoustic speech regions that have the highest potential to spot hostile talking in Algerian Dialect (AD) and German language (GL), we designed four anger detection schemes that exploit four distinct types of speech segments, namely: (i) Voiced regions (V), (ii) Non-voiced waves (N), (iii) Transitions (T) and (iv) the Entire speech waveform (E). These schemes use statistical measures derived from Mel frequency cepstral coefficients and apply support vector machines to detect angry emotion. For both languages, the positive class is composed of vocal expressions that disclose angry mood, while the negative class comprises expressions that convey non-anger temper. The assessment of these schemes shows that speech regions do not exhibit the same performance in detecting hostile speech in AD and GL. The performance measures we found vary with the changes in the number of acoustic attributes and the duration of regions. To offer a comprehensive assessment of all parameters involved in anger detection, we ranked these detectors using a methodology based on weightings of performance indicators. We have found out that the scheme based on V regions is more appropriate for detecting hostile talking in AD. Schemes based on T, N and E waves came in second, third and fourth position, respectively. However, the results for GL revealed that; the detectors based on T, V, E, and N waves were ranked first, second, third, and fourth, respectively.