F0 Parameterization of Glottalized Tones in HMM-Based Speech Synthesis for Hanoi Vietnamese

Abstract

SUMMARYA conventional HMM-based speech synthesis system for Hanoi Vietnamese often suffers from hoarse quality due to incomplete F0 parameterization of glottalized tones. Since estimating F0 from glottalized waveform is rather problematic for usual F0 extractors, we propose a pitch marking algorithm where pitch marks are propagated from regular regions of a speech signal to glottalized ones, from which complete F0 contours for the glottalized tones are derived. The proposed F0 parameterization scheme was confirmed to significantly reduce the hoarseness whilst slightly improving the tone naturalness of synthetic speech by both objective and listening tests. The pitch marking algorithm works as a refinement step based on the results of an F0 extractor. Therefore, the proposed scheme can be combined with any F0 extractor.

Evaluation of speaker-dependent and average-voice Vietnamese statistical speech synthesis systems

“…For glottalized tones such as Broken and Drop tones ("Thanh ngã" and "Thanh nặng" in Vietnamese) and for some creaky voices, particularly those of Northern Vietnamese speakers, it is difficult to extract complete and accurate F0 contours from speech signal due to large variations of the signal's degree of periodicity. Thus the F0 extraction method proposed in [6] was employed in our system to alleviate this problem. Besides, we used the high-quality speech vocoding method STRAIGHT to extract spectral and aperiodicity measurements from speech signals as described in the Nitech-HTS 2005 system [14].…”

“…A couple of HMM-based Text-to-Speech (TTS) systems for Vietnamese have been developed since 2009 [2], [3]. Latest refinements being made to these systems involved in the integration of syntactic information and intonational tags to improve the overall naturalness of generated prosody [4], [5] or the accurate extraction of pitch contours for glottalized tones to enhance the tonal analysis and synthesis [6]. Although the obtained results are promising, all of the above systems are built using the speaker-dependent approach with a moderate amount of training data of one speaker.…”