Yinghao Aaron Li researches speech synthesis and voice generation. He is the first author of StyleTTS 2 and DMOSpeech 2, and announced a move to Meta Superintelligence Labs after his doctoral work at Columbia.[2][4][6]
Start with StyleTTS 2 implementation →
Career
Li’s repositories publish the StyleTTS family, voice-conversion systems and neural vocoder work. In a public account of his final doctoral project, he identifies an Adobe internship behind earlier DMOSpeech research and announces joining Meta Superintelligence Labs. Columbia’s lab lists him among its alumni.[1][6][7]
Contributions
Generate speaking style rather than require a reference
StyleTTS 2 treats speaking style as a latent variable generated through diffusion. A large pretrained speech model helps discriminate the resulting audio during training, and differentiable duration modeling connects timing with the rest of the system. Li is the first author, working with collaborators at Columbia. The paper’s human-listener comparisons apply to its specified datasets and tests.[2]
Optimize how long speech sounds last
DMOSpeech 2 extends direct optimization of speech metrics to the duration predictor. It treats duration generation as a policy and uses reinforcement learning with speaker similarity and word-error signals. The aim is to improve perceived voice identity and intelligibility while retaining fast synthesis. Li is the first author of the collaborative paper.[4]
Tools, writing & talks
Tools
Start here
StyleTTS 2 implementation
github.comThe original implementation, checkpoints and audio-demo links.
Listen to the linked samples, then read the style-diffusion explanation before installing. Running inference needs the repository’s dependencies and checkpoints; training requires speech data and considerably more compute. Read the code’s licence and model conditions before commercial reuse.
Research author and maintainer[3]
Writing, research & talks
- StyleTTS 2arxiv.org
Style diffusion and speech-model adversarial training for text-to-speech.
First author[2] - DMOSpeech 2ojs.aaai.org
Reinforcement learning for duration prediction in metric-optimized speech synthesis.
First author[4]