Applying Random Forest classification to diagnose autism using acoustical voice-quality parameters during lexical tone production

Abstract

Atypical voice quality has been reported among children with autism spectrum disorder (ASD). Yet, it is unclear which acoustic parameters played a crucial role in discriminating the voice of children with ASD from that of their typically developing (TD) peers, especially those who speak a tone language. The current study carried out a preliminary investigation of voice quality in Mandarin-speaking children with ASD using multidimensional acoustic parameters, in an effort to seek the most robust cues using the Random Forest classification. Twenty Mandarin-speaking children with ASD and twenty age-matched TD children participated in the lexical tone production using a picture-naming task. Acoustic parameters included in this study were time-domain parameters: fundamental frequency (F0), the range of F0, the strength of excitation; spectral parameters: H1*-H2*, H2*-H4*, H1*-A1*, H1*-A2*, H1*-A3*; and signal aperiodicity parameters: cepstral peak prominence, harmonic-to-noise ratio, subharmonic-to-harmonic ratio, jitter, and shimmer. Results showed that except for HNR and F0 range, group differences (ASD vs. TD) were found in the other 11 parameters. Additionally, a 78.5% accuracy rate was obtained for classification analysis between voice-quality features of children with and without ASD, with shimmer and jitter as robust parameters. These results indicated that Mandarin-speaking children with ASD tended to overexert and overstrained their voices. Especially for Tone 3 production, they notably exhibited a higher F0 with a less creaky voice, losing the typical voice-quality feature of T3. Although no voice disorders were detected among Mandarin-speaking children with ASD, voice quality has the potential and supplementary value for diagnosing ASD.