I'm about to start new projects in voice synthesis. Which approaches stand out for outputs closest to human speech? For example, are GAN-based models more efficient, or are autoregressive architectures? In terms of realism and naturalness, which methods are preferred? I'd like to benefit from users' experiences.
What are the best voice synthesis methods?
👁️ 4 views💬 1 replies❤️ 0 likes
1 Replies
I recently did some voice synthesis experiments for a YouTube video, and the results were truly striking. At first, I started with GAN-based approaches because I thought I could just "mess around" with audio data to get realistic outputs. But the results started to disappoint me—GANs tend to lose details in frequencies, making the voices sound "synthetic," like a robot speaking through a human mouth.
Then I switched to autoregressive models (like Meshed-TTS and VITS), and the improvement was huge. The voices sounded much more natural—you could easily mimic a podcast host’s voice perfectly. As users have said, with the right data, the human ear can barely tell the difference. Sure, training these models takes longer, but the results are definitely worth it. Plus, I noticed that these autoregressive models nail emotional intonation way better than GANs ever did. So if you're after realism, based on my experience, this is the way to go.