An Age Adapting Electrolarynx - A Feasibility Study

摘要

We propose a mathematical model for voice aging that could be used in the design of an age-adapting Electrolarynx. Voice data from public figures, at the ages of 30, 40, 50 and 60 years old, were acquired from a YouTube corpus. The voice processing consisted of an extraction of 70 Mel-Frequency Cepstral Coefficients (MFCCs) and a computation of their statistical features. ANOVA F-tests were used to determine which of these features change with age. Significant differences between age groups were found only for the first 40 MFCCs. The aging model was then constructed using non-linear regression and an averaged quadratic polynomial fit on these coefficients. Model age-adapted voices were reconstructed from the young dataset speakers' voices and compared to their voices at older ages. The model was validated by the correlation between speakers' MFCCs at older ages and the model-aged MFCCs. The average correlation results were in the range of 0.62 to 0.93. The results imply that the first 40 MFCCs are more susceptible to age related changes and that the proposed model has the potential to enhance the Electrolarynx by providing age adaptation as the speaker grows older.