Reproduce “EmoSpeech: Guiding FastSpeech2 Towards Emotional Text to Speech” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Daria Diatlova, Vitaly Shutov
Researchers at VK, deepvk developed EmoSpeech, an extension of the FastSpeech2 model for emotional Text-to-Speech (ETTS) synthesis that achieves high-quality, emotionally expressive speech while maintaining fast inference. The model achieved an MOS of 4.37 and emotion recognition accuracy of 0.83, surpassing previous FastSpeech2-based ETTS implementations without increasing latency.
Key insights
- Integrating specific architectural modules like eGeMAPS, CLN, CCA, and adversarial training can significantly enhance emotional expressiveness and quality in FastSpeech2 without compromising inference speed.
- Dynamic attention mechanisms, such as Conditional Cross-Attention, are effective for modeling nuanced emotional prosody by allowing the model to emphasize different parts of an utterance based on target emotion.
- Adversarial training helps maintain speech quality in heavily conditioned non-autoregressive TTS models, improving naturalness and overall user perception.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.