Pronunciation modelling in end-to-end text-to-speech synthesis
Sequence-to-sequence (S2S) models in text-to-speech synthesis (TTS) can achieve high-quality naturalness scores without extensive processing of text-input. Since S2S models have been proposed in multiple aspects of the TTS pipeline, the field has focused on embedding the pipeline toward End-to-End (E2E-) TTS where a waveform is predicted directly from a sequence of text or phone characters. Early work on E2ETTS in English, such as Char2Wav  and Tacotron , suggested that phonetisation (lexicon-lookup and/or G2P modelling) could be implicitly learnt in a text-encoder during training. The benefits of a learned text encoding include improved modelling of phonetic context, which make contextual linguistic features traditionally used in TTS pipelines redundant . Subsequent work on E2E-TTS has since shown similar naturalness scores with text- or phone-input (e.g. as in ). Successful modelling of phonetic context has led some to question the benefit of using phone- instead of text-input altogether (see ). The use of text-input brings into question the value of the pronunciation lexicon in E2E-TTS. Without phone-input, a S2S encoder learns an implicit grapheme-tophoneme (G2P) model from text-audio pairs during training. With common datasets for E2E-TTS in English, I simulated implicit G2P models, finding increased error rates compared to a traditional, lexicon-based G2P model. Ultimately, successful G2P generalisation is difficult for some words (e.g. foreign words and proper names) since the knowledge to disambiguate their pronunciations may not be provided by the local grapheme context and may require knowledge beyond that contained in sentence-level text-audio sequences. When test stimuli were selected according to G2P difficulty, increased mispronunciations in E2E-TTS with text-input were observed. Following the proposed benefits of subword decomposition in S2S modelling in other language tasks (e.g. neural machine translation), the effects of morphological decomposition were investigated on pronunciation modelling. Learning of the French post-lexical phenomenon liaison was also evaluated. With the goal of an inexpensive, large-scale evaluation of pronunciation modelling, the reliability of automatic speech recognition (ASR) to measure TTS intelligibility was investigated. A re-evaluation of 6 years of results from the Blizzard Challenge was conducted. ASR reliably found similar significant differences between systems as paid listeners in controlled conditions in English. An analysis of transcriptions for words exhibiting difficult-to-predict G2P relations was also conducted. The E2E-ASR Transformer model used was found to be unreliable in its transcription of difficult G2P relations due to homophonic transcription and incorrect transcription of words with difficult G2P relations. A further evaluation of representation mixing in Tacotron finds pronunciation correction is possible when mixing text- and phone-inputs. The thesis concludes that there is still a place for the pronunciation lexicon in E2E-TTS as a pronunciation guide since it can provide assurances that G2P generalisation cannot.