Sub-Phonetic Modeling For Capturing Pronunciation Variations For Conversational Speech Synthesis

Abstract

In this paper we address the issue of pronunciation modeling for conversational speech synthesis. We experiment with two different HMM topologies (fully connected state model and forward connected state model) for sub-phonetic modeling to capture the deletion and insertion of sub-phonetic states during speech production process. We show that the experimented HMM topologies have higher log likelihood than the traditional 5-state sequential model. We also study the first and second mentions of content words and their influence on the pronunciation variation. Finally we report phone recognition experiments using the modified HMM topologies.

Significance of early tagged contextual graphemes in grapheme based speech synthesis and recognition systems

“…These are compared against the baseline phone based unit selection voice. Segmentation of the database in terms of graphemes and trigraphemes is automatically done using the EHMM labeller [8] in FESTVOX. For the grapheme based system, dictionary representation of each word is in terms of its graphemes, e.g.…”

Data-driven phrasing for speech synthesis in low-resource languages

“…We force align the speech with its transcription using an HMM tool [7]. This tool allows us to bootstrap alignments just from speech and its transcripts for any language.…”

Probabilistic Speaker Pronunciation Adaptation for Spontaneous Speech Synthesis Using Linguistic Features

“…While early work has mostly concentrated on using phonological rules extracted from data to create alternative pronunciations [17,8], most recent techniques are machine learning approaches. Notably, decision trees [7,18], random forests [6], neural networks [5,10], hidden Markov models [16], and CRFs [10] have been investigated. In [18], decision trees and statistical contextual rules are even combined.…”