The dependencies are far too large for use on embedded systems

[AdamWelch1]

If the purpose is to develop TTS that can be run on resource-constrained embedded systems, then you really need to do something about all the dependencies that are pulled in by the setup script. I was trying to benchmark this on a system that only has a 16Gb disk and this eats up almost all of the remaining free space.

As it stands, this is utterly unusable on the very systems it's supposed to be used on.

[gooofy]

:) that is actually what https://github.com/gooofy/zerovox.cpp.git was for - however, given the rapid progress other TTS projects are making at this point I am questioning whether zerovox still has a future. For me, it was a very interesting adventure and tought me a lot about building and training AI systems - however I am very tempted to shut down this project and this point as I suspect there are more modern, better alternatives out there.

[AdamWelch1]

Sorry for such a late reply, I'm usually logged into my work account on Github and haven't logged into my personal account in like... 6 months.

Anyway, I'm not so sure that you should give up, but that's just the opinion of one guy. I work as an embedded software engineer and one big problem that I still see in TTS is that there is still very poor support for embedded systems using older hardware. I work on an older "embedded" system (it's really an x86 desktop machine shoved into a custom box, basically) that has a dual core Intel Atom processor, an integrated GPU, 16Gb of flash memory storage, and 2 Gb of RAM.

The TTS stuff coming out today for embedded systems is still machine learning "AI" based and they seem to still want to run on a GPU? Or perhaps they're using the CPU to execute the tensors, I can't remember. What I do know is that most of them take a ridiculous amount of disk space (mainly python dependencies, model files are usually under 1Gb) and don't run fast enough on the hardware I'm using to be used in real-time applications (for example, GPS navigation while driving).

The best open-source one I've found that runs almost fast enough to be used for on-the-fly generation on my target hardware is Piper, but it takes a variable amount of time depending on the length of the output audio it needs to generate, so in some cases it can take several seconds before outputting the TTS audio.

So far, I've only found 2-3 commercial TTS products that work on my target hardware fast enough to generate TTS audio on-the-fly in real time. There are no open-source alternatives to these products that run well on older hardware. Not a single one that doesn't sound absolutely terrible (espeak-ng works, but sounds awful).

I'm at risk of writing a book here, so I guess my point is that there is still a niche area in TTS that doesn't have any good open-source solution. Everyone that is making TTS for embedded systems is still using neural networks that are too heavy to run on older embedded hardware. If you are interested in still working on TTS, that would be an awesome direction for you to explore. I would definitely be very interested in such a thing.