Are there any free/open-source TTS options out there that are on the same level as Google Cloud’s? I tried a lot of free ones, but they are absolutely awful and still sound like my Amiga did 30 years ago. With LLMs being available as open source, I am hoping there’s also a good TTS offering I just haven’t found yet.
Balabolka was/is my go to for TTS. It creates audio files as well for later if you need. Used it to make plenty of audio books in the past.
Festival – not cutting edge – will definitely be better than your Amiga, and can handle long text. Last time I set it up, IIRC I wanted some voices generated by Tokyo University or something, which took some setting up. It’ll probably be packaged in your Linux distro.
You can listen to a demo here.
https://www.cstr.ed.ac.uk/projects/festival/onlinedemo.html
It’s not LLM-based.
For short snippets, offline, one can use Tortoise TTS – which is LLM based. But it’s slow and can only generate clips of a limited length. Whether it’s reasonable for you will depend a lot on your application. It will let one clone – or make a voice sounding more-or-less similar – a voice using some sound samples from them speaking.
https://github.com/neonbjb/tortoise-tts
Examples at:
https://nonint.com/static/tortoise_v2_examples.html
I haven’t used Google’s, but I’d assume, given that Google is paying people to work on it full time, that whatever they’ve done probably sounds nicer. But, then not open source, so…shrugs
Ah, I looked at Tortoise, but I do not have an nVidia GPU, so I couldn’t try it. Festival I tried and the results were bad. Not so much for the voice, but for intonation and pronunciation.
Ah, I looked at Tortoise, but I do not have an nVidia GPU, so I couldn’t try it.
I use it on an AMD GPU.
EDIT: Wait, let me make sure. I was using an Nvidia GPU for a while and switched to AMD.
EDIT2: Oh, yeah, it uses transformers, and that doesn’t work on rocm presently, IIRC.
I would say Elevenlabs is the best but unfortunately not free.
If you need it for a short while it might be worth it.
I tried Piper with different models, and a couple of FOSS alternatives but the output quality was definitely subpar.
I would say soon we will have good FOSS models, but for the time being that’s not the case.
https://github.com/rsxdalv/tts-generation-webui and https://github.com/gitmylo/audio-webui. I use them all the time. Taking a sample of 10s i get amazing results.
Cool, I’ll give those a try!
Piper is my choice. Very easy to use from the command line, fairly good sounding voices. Prior to that, for years (decades?) I used espeak-ng, had a very robotic voice but articulated almost everything very clearly, and I got used to it so didn’t actually mind.
Espeak doesn’t get better, but nor does it get worse
Came here to recommend Piper. It’s an excellent TTS engine.
Have you tried Piper?
Yes, but if you compare it to https://cloud.google.com/text-to-speech?hl=en (scroll down a bit and you can try it) and the Neural2 model, it sounds like shit. I mean, it’s great to see that there are efforts, but it just pales in comparison.
