Acoustic analysis and recognition of whispered speech

Vocal tract resonances in speech, singing, and playing musical instruments

“…Thus, for in vitro measurements, the first resonance frequency of a cylindrical tube tends to increase with glottal P E R S P E C T I V E aperture and open quotient (Barney et al, 2007). As for in vivo measurements, first resonances have been shown to fall at higher frequencies for whispering than for normal speech, for the same vocal tract shape (the same vowel articulation) Emanuel, 1984a, 1984b;Matsuda and Kasuya, 1999;Itoh et al, 2002;Swerdlin et al, 2008). This can be explained by the larger aperture of the glottis in that mode of phonation (Solomon et al, 1989).…”

“…However, as with the normal voice, the source function is still not known, so the frequencies of the formants may not coincide precisely with those of the resonance. The main drawbacks of these methods are that glottal aperture may be larger for these modes of phonation than for normal speech (discussed later), giving rise to an increase of the first vocal tract resonance frequencies for a similar tract configuration (Matsuda and Kasuya, 1999;Itoh et al, 2002), and that articulation may change from normal to whispered or creak phonations Emanuel, 1984a, 1984b), changing the resonance characteristics as well.…”

“…This area is larger than that used in normal speech, but small compared to that used in breathing. Airflow through it is turbulent, and has a mix of all frequencies (Itoh et al, 2002;Matsuda and Kasuya, 1999). The creak voice or vocal fry has an oscillating glottis, but the variations are not periodic (Hollien and Michel, 1968).…”

An Algorithm for Formant Estimation of Whispered Speech

“…Former studies show that without fundamental frequency, formant estimation becomes prominent in its analysis and recognition [2,3]. In real-world environments where background noise is present, the signal-to-noise (SNR) of whispered speech is lower [4].…”

A Primary Research on Gabor Tensor Sparse Features Representation for Whispered Speech Recognition

“…However, there are few researches on whispered speech. Reference [3] recognizes Japanese whispered speech by using Mel-frequency cepstral coefficient (MFCC) feature and Hidden Markov Models (HMM). It finally has a recognition rate of 68% which can be increased by 10% with maximum likelihood linear regression (MLLR) adaptive training approach.…”