Towards vocal-behaviour and vocal-health assessment using distributions of acoustic parameters

Voice disorders at different levels are affecting those professional categories that make use of voice in a sustained way and for prolonged periods of time, the so-called occupational voice users. In-field voice monitoring is needed to investigate voice behaviour and vocal health status during everyday activities and to highlight work-related risk factors. The overall aim of this thesis is to contribute to the identification of tools, procedures and requirements related to the voice acoustic analysis as objective measure to prevent voice disorders, but also to assess them and furnish proof of outcomes during voice therapy. The first part of this thesis includes studies on vocal-load related parameters. Experiments were performed both in-field and in laboratory. A one-school year longitudinal study of teachers’ voice use during working hours was performed in high school classrooms using a voice analyzer equipped with a contact sensor; further measurements took place in the semi-anechoic and reverberant rooms of the National Institute of Metrological Research (I.N.Ri.M.) in Torino (Italy) for investigating the effects of very low and excessive reverberation in speech intensity, using both microphones in air and contact sensors. Within this framework, the contributions of the sound pressure level (SPL) uncertainty estimation using different devices were also assessed with proper experiments. Teachers adjusted their voice significantly with noise and reverberation, both at the beginning and at the end of the school year. Moreover, teachers who worked in the worst acoustic conditions showed higher SPLs and a worse vocal health status at the end of the school year. The minimum value of speech SPL was found for teachers in classrooms with a reverberation time of about 0.8 s. Participants involved into the in-laboratory experiments significantly increased their speech intensity of about 2.0 dB in the semi-anechoic room compared with the reverberant room, when describing a map. Such results are related to the speech monitorings performed with the vocal analyzer, whose uncertainty estimation for SPL differences resulted of about 1 dB. The second part of this thesis was addressed to vocal health and voice quality assessment using different speech materials and devices. Experiments were performed in clinics, in collaboration with the Department of Surgical Sciences of Università di Torino (Italy) and the Department of Clinical Science, Intervention and Technology of Karolinska Institutet in Stockholm (Sweden). Individual distributions of Cepstral Peak Prominence Smoothed (CPPS) from voluntary patients and control subjects were investigated in sustained vowels, reading, free speech and excerpted vowels from continuous speech, which were acquired with microphones in air and contact sensors. The main influence quantities of the estimated cepstral parameters were also identified, which are the fundamental frequency of the vocalization and the broadband noise superimposed to the signal. In addition, the reliability of CPPS estimation with respect to the frequency content of the vocal spectrum was evaluated, which is mainly dependent on the bandwidth of the measuring chain used to acquire the vocal signal. Regarding the speech materials acquired with the microphone in air, the 5th percentile resulted the best statistic for CPPS distributions that can discriminate healthy and unhealthy voices in sustained vowels, while the 95th percentile was the best in both reading and free speech tasks. The discrimination thresholds were 15 dB (95\% Confidence Interval, CI, of 0.7 dB) and 18 dB (95\% CI of 0.6 dB), respectively, where lower values indicate a high probability to have unhealthy voice. Preliminary outcomes on excerpted vowels from continuous speech stated that a CPPS mean value lower than 14 dB designates pathological voices. CPPS distributions were also effective as proof of outcomes after interventions, e.g. voice therapy and phonosurgery. Concerning the speech materials acquired with the electret contact sensor, a reasonable discrimination power was only obtained in the case of sustained vowel, where the standard deviation of CPPS distribution higher than 1.1 dB (95\% CI of 0.2 dB) indicates a high probability to have unhealthy voice. Further results indicated that a reliable estimation of CPPS parameters is obtained provided that the frequency content of the spectrum is not lower than 5 kHz: such outcome provides a guideline on the bandwidth of the measuring chain used to acquire the vocal signal.