Analysis of Short-Time Magnitude Spectra for Improving Intelligibility Assessment of Dysarthric Speech.

The discriminating information about dysarthria exists at the fine level in the short-time Fourier transform magnitude spectra (STFT-MS). To capture the discriminating information present in STFT-MS using Mel-frequency cepstral coefficients (MFCCs), this paper firstly studied the role of increasing the size of Mel-filterbank and inverse Mel-filterbank. A novel feature extraction technique is then proposed for the assessment of dysarthria using speech. In the proposed approach, the STFT-MS is processed through an accumulator (digital integrator) to capture the spectral dynamics (SD). The accumulator output over a frequency range is a growing or decaying function of frequency depending on the peaks and valleys present within that region due to the pitch and resonance structure of the vocal tract system. The SD over a band of frequency is computed from the accumulator output by finding non-local differences between frequency points placed linearly in non-overlapping mode. The SDs are logarithmically compressed (LCSD) to normalize the magnitude of SD computed in different frequency regions. The LCSD represents (M) dimensional feature vector when it is computed at M linearly spaced frequency points. The i-vector-based dysarthric-level assessment system on the universal access speech reported in this study shows that the performance of the MFCC feature improves significantly by increasing the Mel-filterbank size. The MFCCs computed from the inverse Mel-filterbank (IMFCCs) contain additional information to MFCCs. The use of LCSD provides improved performance than MFCCs and IMFFCs and is also provides additional to MFCCs. [ABSTRACT FROM AUTHOR]