Recent advances in Text-to-Speech (TTS) synthesis have enabled the generation of highly intelligible and natural-sounding speech. However, most conventional TTS systems still produce neutral or reading-style speech and therefore lack the expressiveness required for applications such as audiobook narration, virtual assistants, and human–machine interaction. While expressive speech synthesis has traditionally focused on prosodic variation, growing evidence suggests that pronunciation variation also plays an important role in conveying speaking style and emotion.
This study investigates the contribution of pronunciation variation to expressive speech synthesis in French and explores whether deep neural networks can automatically transform canonical pronunciations into speech-style-dependent pronunciations. Particular attention is given to pronunciation phenomena characteristic of expressive French speech, including the realization or omission of the mute e (schwa-like vowel) and variations in pause realization between words.
The research formulates the problem as a grapheme-to-phoneme (G2P) conversion task and develops a series of speech-style-aware pronunciation models based on a sequence-to-sequence encoder–decoder recurrent neural network (RNN) architecture implemented in OpenNMT. Several modeling strategies are evaluated, including the use of shorter phoneme sequences, explicit word-boundary information for pause prediction, and transfer learning through model pre-training. Model performance is assessed using the phoneme error rate (PER) on six neutral-to-emotion pronunciation conversion tasks in a French expressive speech corpus.
The experimental results show that deep neural network-based G2P conversion can successfully model pronunciation variation associated with expressive speech and consistently improves pronunciation prediction accuracy. The findings demonstrate that pronunciation variation should be considered alongside prosodic modeling in expressive speech synthesis and provide a practical framework for speech-style-aware grapheme-to-phoneme conversion in French.
- Citar trabajo
- Ruoxiao Yang (Autor), 2020, Speech-Style-Aware Grapheme-to-Phoneme (G2P) Conversion for Expressive Speech Synthesis, Múnich, GRIN Verlag, https://www.grin.com/document/1742322
