@KaraLG84 Interestingly, I did use DECtalk first. It butchered lots of other stuff, but got numbers fine.

Ah damn. Win some lose some.

I could just explicitly convert any weird pronunciations to phonemes and use DECtalk anyway. I might still do that.

@KaraLG84 Am I the only one who likes the TTS voices that sound a little bit synthetic? I don't like to be kept guessing. I think voices like SAPI 5 Tom and AT&T Crystal struck a good balance. If it's good enough to use on trains, it's good enough for a bot.

@Mayowa What's wrong with 1.67? That's what the text actually said. As for the rest, yeah, it's broken.

@sclower Interestingly, the V1.0 API is perfectly fine with it. V2 and especially V3 are all kinds of weird.

This is the thing, if they were even a little transparent about what exactly they were doing under the hood I could come up with at least an educated guess as to why it struggle so much with numbers and other symbols, but I have no idea. I'm pretty sure the model they're using is some kind of LLM architecture, similar to many recently released open-source TTS models, and those are notorious for struggling with numbers and other symbols without text normalization being done. I'll be honest, Eleven is actually behind the competition from the likes of Amazon and Microsoft, and many of the open source options not only sound just as good, but are much more customizable as well.

@ZBennoui @sclower Fortunately, we will soon have a dedicated, in-house server to run stuff on, so the possibility of using reasonable local models will exist once a decent GPU is added.

@sclower Well you'll certainly have a lot of options to pick from. Something that could be fun and potentially useful is to try fine-tuning a model on the type of content you're going to be generating, performance will likely improve drastically. These LLM based models are much easier to find tune than traditional ML TTS, I was able to train an Orpheus TTS model on my own voice in about ten minutes.