Probablistic modelling of F0 in unvoiced regions in HMM based speech synthesis

Abstract

HMM based synthesis has attracted great interest due to its compact and flexible modelling of spectral and prosodic parameters. In this approach, short term spectra, fundamental frequency (F0) and duration are simultaneously modelled by multi-stream HMMs. However, since F0 values in unvoiced regions are normally considered as undefined, it is difficult to use standard HMMs for F0 modelling. The currently preferred solution to this is to use a multi-space distribution HMM (MSDHMM) in which discrete distributions are used for modelling the voiced/unvoiced decision and continuous Gaussian distributions are used for modelling the F0 values within the voiced regions. However, the assumption of undefined unvoiced F0 regions and the special structure of the MSDHMM lead to limitations in the accurate modelling of F0 patterns. In this paper an alternative is explored whereby unvoiced F0 values are assumed to exist and are modelled within the standard HMM framework using a globally tied distribution (GTD). Subjective evaluations show that these regular HMMs with GTD can produce significant improvements in the naturalness of the synthesised speech compared to the MSDHMM, and furthermore, the method is insensitive to the exact method used for unvoiced F0 generation.

Joint modelling of voicing label and continuous F0 for HMM based speech synthesis

“…By generating real F0 values for unvoiced regions and assuming hidden voicing labels, the Continuous F0 model with Globally Tied Distribution [6], CF-GTD in figure 1(b), is obtained. The state output distribution can be expressed as…”

“…Experiments have shown that it can greatly reduce the F0 trajectory modelling error, and consequently improve the naturalness of the synthesised speech [6,7]. However, due to hidden voicing labels, voicing classification only relies on the statistical difference between the globally tied unvoiced component and the state specific voiced component.…”

“…Although different unvoiced F0 value generation approaches have in the past been shown to give similar performance [6,7], all methods have been static. To reflect the randomness property of unvoiced F0, the dynamic random generation approach described above was investigated.…”

“…Here, continuous F0 is assumed to exist in unvoiced regions and there have been a number of modelling approaches along this line. In [6], random F0 values are used in unvoiced regions and voicing labels are assumed to be hidden. A Gaussian mixture model (GMM) is employed, where unvoiced Gaussian components are globally tied so that the statistical difference between voiced and unvoiced regions can be modelled.…”

“…In [7], voicing labels are assumed to be observable and modelled in an independent stream. As the voicing labels are explicitly modelled, global tying as defined in [6] is no longer a requirement for distinguishing voiced regions from unvoiced regions. Both approaches have shown significant improvement in the naturalness of synthesised speech compared to the traditional MSDHMM approach.…”