Fundamental frequency tracking in diplophonic voices. (August 2017)

• Record Type: Journal Article
• Authors: Aichinger, P.; Hagmüller, M.; Roesner, I.; Schneider-Stickler, B.; Schoentgen, J.; Pernkopf, F.
• Is Part Of: Biomedical signal processing and control. Volume 37(2017)
• Page Start: 69
• Page End: 81
• Publication Date: 2017-08
• Subjects: AWM audio waveform modeling -- ASA audio subband auto-correlation -- DFT discrete Fourier transformation -- FIR finite impulse response -- fo fundamental frequency -- FS Fourier synthesis -- HMM hidden Markov model -- SPP spectral peak picking; Laryngeal high-speed videos -- Diplophonia -- fo Tracking -- Voice disorders -- Audio signal processing; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2016.10.002
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Diplophonic voices involve two simultaneous fundamental frequencies. Three independent approaches to fundamental frequency tracking were compared. Reference tracks were obtained via cycle boundaries marked in kymograms. Audio waveform modeling based tracks agreed with kymogram based tracks. Audio subband auto-correlation based tracks were wrong for diplophonic voices. Abstract: Background and objectives: Fundamental frequency ( fo ) extraction in disordered voices is a prerequisite for many types of clinical analyses. Special attention must be paid if multiple oscillators with different fo s are active simultaneously. Two independent approaches to fo tracking in diplophonic voices are proposed and compared with a benchmark from the literature. Material and methods: Six samples of sustained phonations were analyzed. High-speed videos were obtained in addition to audio recordings. Video-based fo tracks were obtained from cycle marks that report maximal vocal fold deflection in digital kymograms. Audio waveform modeling based extraction involved candidate tracking, oscillator waveform synthesis and track selection. Audio subband auto-correlation based extraction served as a benchmark. Results and discussion: Promising qualitative and quantitative agreement of audio waveform modeling based estimates with kymogram-based tracks was observed. With reference to the kymogram-based tracks, audio waveform modeling based extraction had a median total error rate of 1.9%, which is an improvement over the benchmark method (17.7%). Conclusion: The results illustrate that fo s of diplophonic voices may be validly obtained from kymogram cycle marks, as well as via audio waveform modeling. The acquisition of two simultaneous fo tracks in diplophonic voices may increase the validity of clinical voice analysis procedures in the future.