Sınırlı veri ile denetimli duygusal konuşma sentezi için derin sinir modellerinin incelenmesi

Human speech serves as a richly expressive medium, conveying a multitude of information beyond mere words, including the speaker's identity, emotions, and different speaking styles. Today's Text-to-Speech (TTS) models, powered by deep learning, have advanced to the point of producing speech that is almost human-like in its naturalness and intelligibility. Nevertheless, there remains considerable ground to cover in the field of expressive speech synthesis when compared to the expressiveness of human speech. Supervised emotional TTS models (ETTS), capable of generating speech across various emotional states, primarily depend on training data annotated with multiple emotions. The creation of such datasets poses significant challenges, resulting in limited availability and typically smaller sizes. In this thesis, we delve into the investigation of various deep neural models aimed at improving the emotion expressivity of supervised ETTS approaches. The models explored in our research include voice conversion models, emotion classifiers, and adversarial training discriminators. We propose two extended ETTS models: the first combines an emotional voice converter into the ETTS model, while the second incorporates two ensembles, each comprising an emotion classifier and a discriminator, into the ETTS architecture. Within the first model, we examine two distinct voice conversion models: MaskCycleGAN and Seq2Seq voice converters. In the second model, two types of emotion expressivity-related features are employed as input to the ensembles: pitch features and emotion embeddings generated by a variational autoencoder. We conducted several evaluation tests to assess the feasibility of the applied components and the overall performance of each proposed model. Our experiments revealed significant improvements by the proposed models over the baseline model.