Method of detection and classification of nasalized vowels based on acoustic features derived from both magnitude and phase spectra

Abstract

The performance of the detection and classi cation of nasalized vowels from the mixture of oral and nasalized vowels largely depend on the acoustic features by which the task is performed. In this thesis, sets of acoustic features are derived from both tha magnitude and phase spectra of the vowels to evaluate the performance. It is shown through detail acoustic analysis on nasalized vowels that, although ad- ditional formants at various frequency locations are introduced, a new resonance in the low frequency region around 250 Hz is introduced and found to remain consistent irrespective of male or female speakers in the modi ed group delay function. By ex- ploiting this fact, which is veri ed on the band limited modi ed group delay function capable of resolving two closely spaced formants occuring in case of nasalized vowels, an acoustic parameter RMGD is derived. Utilizing RMGD, the idea of detecting nasalized vowels and determining the degree of nasality with respect to the adjacent nasal consonants of the vowel is evolved. It is argued that vowel can be nasalized with at least one adjacent nasal consonant even if the nasal consonant is pre-vocalic and the vowel is more nasalized in pre-nasal position than in post-nasal position. It is also found that vowel with nasal consonants on both side do not guarantee to be more nasalized vowel compared to the vowel with one adjacent nasal conso- nant. By utilizing the fact of changing nasality with the number of adjacent nasal consonant, the detection and classi cation of non-nasalized and contextually nasal- ized vowels is formulated as a four class problem and solved based on a threshold or classi er based scheme and found superior in detecting and classifying nasalized vowels than some of the existing methods. Mel-ferequency Cepstral coe cients are widely used features for the detection task. Conventionally, features for detecting and classifying nasalized vowels are derived considering magnitude spectrum only, ignoring the phase spectrum. Exploiting the power spectrum and the group delay function of a band limited vowel, the product spectrum is de ned thus incorporat- ing the information of both magnitude and phase spectra. The product spectrum is then di erentiated with respect to frequency to obtain di erential product spectrum that is argued to provide more noise robustness in the presence of noise. Assuming the noise reduction capability of the autocorrelation sequence power and product spectrum of the band limited autocorrelation sequence of the vowel are developed. Simulation results show that Mel-frequency cepstral coe cients derived from the product spectrum, di erential product spectrum, power and product spectrum of the autocorrelation sequence consitently outperforms the some of the convetional approaches in the task of detecting and classifying nasalized vowels in both clean and di erent noisy conditions.