Objective
Cepstral peak prominence (CPP) is an acoustic measure that provides valuable information about vocal periodicity and is strongly related to the perception of dysphonia. Although CPP is currently recommended for acoustic evaluation of dysphonia, previous research has shown that CPP can be affected by background noise. This is critical, as background noise is often present in clinical and research voice recordings unless recordings are made in a well-controlled condition. This study aims to examine the extent to which background noise affects CPP and identify the threshold at which it can be present without causing a significant change in CPP.
Methods
Audio recordings from 50 dysphonic children with vocal fold nodules (ages 3-10) producing sustained vowels and connected speech will be examined. All data has previously been collected in a quiet environment during clinical evaluation. Different types of background noise (e.g., white noise, multi-talker babble) will be mixed with the speech in MATLAB, creating new files with SNRs ranging from poor (0 dB) to ideal (50 dB) with 0.51 dB increase per timestep. The CPP value will be calculated in Praat for each updated file.
Results
To date, new SNR files have been created for the sustained /a/ production from 20 children (10 female, 10 male) for the white noise condition. 201 files were created per child (SNR ranging 0-50). Preliminary findings indicate that increased noise (lower SNR) leads to greater deviation from baseline CPP, with larger changes seen around 30 dB. For example, at around 50 dB SNR, the average change from baseline was 0.004, while for 30 dB SNR, it was 0.43. Change from baseline increased nonlinearly as the SNR worsened (20 dB = 1.52 change from baseline, 15 dB = 2.61, 10 dB = 3.87, 0 dB = 6.77).
Conclusions
The change in CPP relative to the SNR appears to follow a nonlinear trajectory, and time-series analysis will be conducted on the full dataset, providing information on the inflection point at which CPP changes significantly. The presentation will discuss the full dataset and examine sustained vowels and connected speech across multiple types of background noise.