There's a lot of stuff out there for vocal synthesis. It's a field that has seen research and work for decades. There's a high representation of content from Japanese researchers, it seems. I'm focusing on musical/singing aspects, the sure leader of which is Vocaloid. However, Vocaloid is very much proprietary and entirely closed off. However, there are alternatives:
Then, there's some amount of actual open-source stuff out there as well. The most of it seems to be based off of [HTS], a speech synthesis system that uses Hiden Markov Models and Deep Neural Networks to synthesize audio. [sinsy] is an open source singing synth that builds on top of HTS to generate synthesized singing. It's still under development very recently (as of Dec 2020 they released some significant new features). However, it seems that they stopped publishing source code updates with their latest features, and basically none of their trained voice models are available. Another popular one out there that builds on top of HTS is [open-jtalk] which is like Sinsy but targeted at speech rather than song.
From what I can tell, there really isn't a readily available drop-in open source solution available. [sinsy] is the closest that I've found to that, but as I mention in the article many of their best/most recent features don't seem to be available in their open source release and none (well, all but one crappy one) of their voice models seem to be publicly available. Building something broadly useful in this space for web-synth will almost certainly require a ton of effort and be a massive undertaking in and of itself.
Recently, as in the past 3-5 years, there has been an explosion of activity in neural-network powered audio generation. There has been a lot of research from places like Google invested in very strong TTS for obvious reasons. The strongest and most actively researched member of this space seems to be [wavenet], a generative artificial neural network that produces output audio sample-by-sample rather than using vocoders or anything else like that. It is what powers Google's in-house TTS across their suite of products and software.
The more I read into these things, the more accessible it actually seems to be. There are a lot of high-quality implementations of the research papers created by places like NVIDIA. Training data for the [tts] parts
Even more recently, some very powerful and easy-to-use end-to-end vocal synthesis platforms have been released. Most notable to me is [larynx] and the underlying [glowtts]; they almost commoditize the whole of TTS. 50+ pre-trained voices, etc.
There is a notable exception to much of my assessment above in the form of a single person: [sleepwalking]. His Github has a collection of seemingly high-quality libraries that are still updated and actively used in his [synthesizer-v] software, meaning that they have real-world use and applicability in a popular, high-quality piece of vocal synthesis software.