Low-Resource Speech Synthesis with Speaker-Aware Embedding
Li-Jen Yang, I-Ping Yeh, Jen‐Tzung Chien · 2022 13th International Symposium on Chinese Spoken Language Processing (ISCSLP) · 2022
Speech synthesis has been successfully exploited for mapping from text sequence to speech waveform where high-resource languages have been well studied and learned from a large amount of text-speech paired data in public-domain corpora. However, developing speech synthesis under low-resource languages is challenging for speech communication in local regions since the collection of training data is expensive. In particular, the speaker-aware speech generation under low-resource settings is crucial in real world. Such a problem is increasingly difficult in case of very limited speaker-specific data. This paper presents a speaker-aware speech synthesis under low-resource settings based on an encoder-decoder framework by using transformer. Knowledge transfer is performed by incorporating a speaker-aware embedding through first learning a pretrained transformer from multi-speaker data of a low-populated spoken language and then fine-tuning the transformer to a target speaker with very limited speaker-specific embeddings. Experiments on low-resource Taiwanese speech synthesis are evaluated to show the merit of speaker-aware transformer in terms of Mel cepstral distortion and mean opinion score.