Improving TTS with Corpus-Specific Pronunciation Adaptation

Abstract

Text-to-speech (TTS) systems are built on speech corpora which are labeled with carefully checked and segmented phonemes. However, phoneme sequences generated by automatic grapheme-to-phoneme converters during synthesis are usually inconsistent with those from the corpus, thus leading to poor quality synthetic speech signals. To solve this problem, the present work aims at adapting automatically generated pronunciations to the corpus. The main idea is to train corpusspecific phoneme-to-phoneme conditional random fields with a large set of linguistic, phonological, articulatory and acousticprosodic features. Features are first selected in cross-validation condition, then combined to produce the final best feature set. Pronunciation models are evaluated in terms of phoneme error rate and through perceptual tests. Experiments carried out on a French speech corpus show an improvement in the quality of speech synthesis when pronunciation models are included in the phonetization process. Appart from improving TTS quality, the presented pronunciation adaptation method also brings interesting perspectives in terms of expressive speech synthesis.Experiments were carried out on a French speech corpus dedicated to interactive vocal system TTS. As such, this corpus covers all diphonemes present in French and comprises most

Can We Generate Emotional Pronunciations for Expressive Speech Synthesis?

“…A perceptual study [40] has shown that samples synthesized with the target pronunciation were preferred to those synthesized with the canonical pronunciation. Also, the adaptation of the canonical pronunciation to the voice corpus has shown a clear preference in terms of quality [4]. However, it seems that the generation of spontaneous speech requires some compromises between intelligibility and quality [41].…”

“…The emotional P2P system should fit pretty well with emotional pronunciation, thus increasing the expressivity of output speech samples, but will probably overfit the data. Moreover, if this set-up is not adapted to the voice corpus, then inconsistencies between the corpus used for synthesis and the corpus used for pronunciation remain, lowering the TTS quality [4]. Fig.…”

“…The corresponding set of 60 features presented in Table 2 is inspired from [41]. It has been enriched and adapted to French in [4]. Previous and next words are added in the feature set.…”

“…1b. In the applied method already presented in [4], features are selected separately for each of the four groups of features (linguistic, phonological, articulatory and prosodic) reported in Table 2. For each group of features, three symmetric phoneme window sizes are tested (W 0 : current phoneme only, W 1 : current, previous and next phonemes, W 2 : current, 2 previous and 2 next phonemes).…”

“…In the present article, grapheme sequences are automatically converted into canonical phoneme sequences using a rule-based phonetizer. Starting from a pronunciation adaptation method originally developed to improve the perceived quality in TTS [4], expressive pronunciation models are trained to adapt canonical neutral pronunciations to target emotional pronunciations as transcribed in an emotional pronunciation corpus. Different aspects of the automatic adaptation framework are studied and evaluated at the phonetic level.…”