Speech Synthesis Based on Hidden Markov Models

Abstract

This paper gives a general overview of hidden Markov model (HMM)-based speech synthesis, which has recently been demonstrated to be very effective in synthesizing speech. The main advantage of this approach is its flexibility in changing speaker identities, emotions, and speaking styles. This paper also discusses the relation between the HMM-based approach and the more conventional unit-selection approach that has dominated over the last decades. Finally, advanced techniques for future developments are described. Index Terms-text-to-speech synthesis, hidden Markov model, HMM-based speech synthesis, statistical parametric speech synthesis, HTS The research leading to these results was partly funded by JST CREST

Voice Conversion Using Input-to-Output Highway Networks

“…After applying a weighting matrix W [3] to an input speech parameter sequence x = [x 1 , · · · , x T ] for calculating its static-dynamic speech feature sequence, the DNNs predict a static-dynamic speech feature sequence of the converted speech.ŷ is generated from the static-dynamic features by using the maximum likelihood-based parameter generation algorithm [2]. We define the above speech parameter conversion asŷ = G(x).…”

“…Deep Neural Networks (DNNs) [1] have been used as acoustic models for VC because they can represent the relationship between the input and output speech parameters more accurately than conventional Gaussian mixture models [2]. These acoustic models are trained with training algorithms such as the maximum likelihood criterion [3] and Minimum Generation Error (MGE) criterion [4], [5]. However, the converted speech parameters tend to be oversmoothed, and this phenomenon degrades the quality of the converted speech.…”

Investigation of Using Continuous Representation of Various Linguistic Units in Neural Network Based Text-to-Speech Synthesis

“…The front-end of a TTS system infers the symbolic representation of both segmental and prosodic properties of speech. Then, the back-end acoustic model converts the symbolic intermediate representation into a speech waveform, typically using the unit-selection method [8] or the statistical parametric method [9].…”

“…The acoustic model based on the hidden Markov model (HMM) has dominated the parametric speech synthesis method for decades [9]. However, its decision-tree-based model clustering method may not be able to express complex dependency in the input linguistic representation of text [23].…”

Non-Native Text-to-Speech Preserving Speaker Individuality Based on Partial Correction of Prosodic and Phonetic Characteristics

“…There have been many attempts at developing crosslingual speech synthesis based on statistical voice conversion [9] or Hidden Markov Model (HMM)-based speech synthesis [10]. For example, one-to-many Gaussian Mixture Model (GMM)-based voice conversion can be applied to unsupervised speaker adaptation in cross-lingual speech synthesis [11], [12].…”