The ability to generate high-quality synthetic speech with fine-grained control over intonation and emotion is crucial for broadcast applications such as narration and voice-over production. This paper presents a novel text-to-speech (TTS) workflow designed to generate expressive and user-controllable speech synthesis. By leveraging a one-hot encoded emotional embedding approach through application-aware, manually labeled emotion classes and a fine-tuned neural model, our method allows for precise control over speech tone and intention. We detail the data preparation pipeline, including manual annotation and automatic emotion classification, and describe the implementation of our speech synthesis model using VITS2. Additionally, we introduce an intuitive prosody control mechanism that enables word- and phoneme-level adjustments for pitch and duration. Subjective evaluation results indicate that our synthetic narration achieves parity with professionally recorded voice-overs, with some preference for synthetic speech due to its enhanced customization. This research contributes to the advancement of AI-driven voice production, enabling scalable, high-fidelity, and fine-grained controlled speech synthesis for media and entertainment industries.
(See document for exact affiliation information.)
https://aes2.org/publications/elibrary-page/?id=23020
(457KB)
Click to purchase paper as a non-member or login as an AES member. If your company or school subscribes to the E-Library then switch to the institutional version. If you are not an AES member Join the AES. If you need to check your member status, login to the Member Portal.
