As anybody who has listened to Artisan Radio knows, we do automated voice tracking. It's all done via preprocessing playlist audio - cutting off beginning and end silence for each, normalizing the track and then using Text To Speech to generate the voice track. Finally, the two chunks of audio are concatenated, and all metadata transferred to the new audio. All this is done within a batch file, so I can do hundreds at a time.
It currently works pretty well, but I haven't really done anything to improve the voice tracking mechanism in a while. So I did some research, and found this open source program - kokoro-TTS, that does a much better job. The voices (and there are a lot of them) are improved over what I use, and it seems to do better at pronunciation and expression. Mind you, I haven't tested it out thoroughly yet, I've just run a few examples through it. But they all sounded great.
I found that the default voice in the examples, af_Sarah, was the best (which makes sense). I wanted a female voice so I never checked the male ones (but I will eventually get around to it).
I did try out some of the online AI sites as well, but most of them want money up front (once you get past the try it once stage), and to be honest, I think that kokoro-TTS does a much better job at TTS. If you want more, such as auto generation of text or lyrics, or mixing in music, then you really have no choice but to go to these AI sites.
But I'm happy.
Here's an update.
I was able to easily modify my audio processing batch file to use kokoro-tts and generate voice tracking. After running through 130 files, I then went through and sampled the results.
There were a few mispronunciations, but overall it sounds much better than what I had. You can put phonemes into the text to be converted, but they're non standard (apparently the kokoro software uses another open source project, misaki, for this). That's one advantage that balabolka (the TTS software I was using) has over kokoro - you can create a dictionary in that software and it will automatically use that to correct any mispronunciations.