Singing Synthesis Beyond Human-Level Naturalness: Not What You Think

Explore a groundbreaking presentation on singing synthesis that challenges conventional notions of human-level naturalness in vocal synthesis research. Delve into the intriguing findings of a system that surpasses raw recordings in comparative mean opinion score tests, while examining the subtle yet crucial distinctions between true human parity and competitive ratings. Unpack the complexities of subjective quality evaluation, analyze the unique challenges posed by singing versus speech synthesis, and consider the implications for future singing synthesis system designs. Learn from Kanru Hua, founder of Dreamtonics and developer of Synthesizer V, as he shares insights on bridging speech signal processing algorithms with cutting-edge generative models and addressing production challenges in deploying neural networks for audio processing.