And yeah, some models are worse than others. I would avoid models trained on voices from tv shows or games, because they usually have a very limited range and break a lot with natural speak. Voice models trained off of youtubers, livestreamers etc. tend to work a lot better.

After you download it, unzip it with 7zip to your drive, and run run-web-gui.bat, and it will open a terminal window which says 'if this is the first time running, it maybe take 20 seconds to start', but its written in Chinese, but now you know what it means. Then it should start.

Also, the problem with Japanese models, is that the training data usually doesnt contain any audio of the person speaking in English, so every time you use the model and make a sound that wasnt in the training data (typical English sounds, especially with the letter 'r'), it has to retrieve sound from the pretrain instead, which will sound different from the intended voice and thus you end up with a voice that doesnt sound as good as in the sample.

If you could however find a similar enough model that was made with English training data, you could merge that model with this one to 'fill in the blanks' and try out different ratios and end up with something that might sound really good. You could also make an rvc model from your real voice, and then merge it with the Japanese model at a low ratio, where its maybe only 5 or 10% your own voice, and get a big bump in quality.

The last software I tried was Vonovox, which works decently well, and supports SPIN. Im going to download the new RVC and test it out.