Effect of voice quality on frequency-warped modeling of vowel spectra

Abstract

The perceptual accuracy of an all-pole representation of the spectral envelope of voiced sounds may be enhanced by the use of frequency-scale warping prior to LP modeling. For the representation of harmonic amplitudes in the sinusoidal coding of voiced sounds, the effectiveness of frequency warping was shown to depend on the underlying signal spectral shape as determined by phoneme quality. In this paper, the previous work is extended to the other important dimension of spectral shape variation, namely voice quality. The influence of voice quality attributes on the perceived modeling error in frequency-warped LP modeling of the spectral envelope is investigated through subjective and objective measures applied to synthetic and natural steady sounds. Experimental results are presented that demonstrate the feasibility and advantage of adapting the warping function to the signal spectral envelope in the context of a sinusoidal speech coding scheme.

Introduction

A successful low bit rate speech coding algorithm requires a good model for the speech signal together with an effective parameter quantization algorithm. A popular model for low bit rate speech coding has been the sinusoidal model, an important example of which is the multiband excitation (MBE) model (Griffin and Lim, 1988). In the case of voiced speech, the parameters of the model are the fundamental frequency (or pitch), and the amplitudes and phases of the harmonics. The harmonic amplitudes represent the product of the source excitation and vocal tract spectra. At low bit rates, estimated phases are usually dispensed with, and the accurate representation of the pitch and harmonic amplitudes becomes critical to the perceptual quality of decoded speech. Steady vowel sounds are particularly sensitive to harmonic amplitude representation errors.

The quantization of harmonic amplitudes is most demanding on the bit allotment in sinusoidal coding, and methods for efficient quantization have been an important topic of research. A widely used method for the quantization of the amplitudes of the harmonics is based on the modeling of a spectral envelope fitted to the harmonic peaks (MacAulay and Quatieri, 1995). The spectral amplitudes are then reconstructed from samples of the modeled spectral envelope at harmonic frequencies. Representing the spectral envelope by the coefficients of an all-pole filter enables the use of one of the many efficient quantization methods available in the speech coding literature. The order of the all-pole model has a significant effect on the accuracy of the modeled spectral amplitudes. While the all-pole representation of the spectral envelope is expected to capture local resonances accurately, capturing additional features such as overall spectral tilt and spectral zeros due to nasality typically lead to an increase in the number of poles required for an adequate approximation. Further, for similar spectral envelope, low pitched sounds require higher LP model order for similar perceived quality levels (Champion et al., 1994, Rao and Patwardhan, 2005).

In the interest of achieving low bit rates, however, it is necessary to keep model order as low as possible. Frequency-scale warping before all-pole modeling of the spectral envelope is a widely used method to improve the perceptual accuracy of modeling for a given model order. Frequency-scale warping leads to a more accurate representation of the low frequency spectrum at the cost of increased errors in the high frequency region. Although perceptual scales such as the Bark-scale and its variants have been widely used in LP modeling of speech spectra, our recent work (Rao and Patwardhan, 2005) on synthetically generated steady vowel sounds (using fixed excitation source parameters) indicated that the performance of frequency warping depended to a great extent on the nature of the underlying sound spectrum. It was observed that front vowels such as [ei] and [i] in fact were better modeled without Bark-scale warping. This was explained by the low frequency first formant structure of these vowels which fails to mask high frequency distortion adequately. In a subjective experiment using natural speech, it was found that the comparative behaviour of modeling error under different warping conditions in the case of the non-front vowels was inconsistent across speakers indicating a further dependence on speaker voice quality as reflected in the overall spectral envelope.

In the present study we attempt to extend our previous work by exploring aspects of voice quality that could influence the spectrum modeling error. The influence of voice quality on perceived modeling error in frequency-warped all-pole modeling of the spectral envelope is studied via the framework of MBE model analysis–synthesis.

While the overall spectral envelope for voiced speech is determined by the glottal source waveform, vocal tract transfer function and lip radiation, it is the glottal source waveform that most directly affects the relative strengths of the low and high frequency harmonics. The scope of the present study is restricted to variations in laryngeal voice qualities for they are closely linked to gross differences in spectral envelope. An investigation based on subjective and objective experimental evaluation is presented. Finally, the applicability of the results to improving speech quality in a low bit rate sinusoidal coder is discussed.