someone on a local llm subreddit asked which text to speech engines people actually run locally and i fell into the whole thread comparing the two big names against my own setup. omnivoice is the community favorite, the most named daily driver. people say it clones voices really well and handles accents great but it struggles with performance and gives you very few ways to steer emotion. benchmarks back that up, it runs up to 40x real time on a 4090 with about two gigs of vram and even runs near real time on an integrated gpu way weaker than my desktop card. it comes from the kaldi team, ten thousand stars, over a million downloads, so not some random repo. the catch is licensing, the code is apache but the pretrained weights are non-commercial only. perfect for personal use, dead end the second you want to sell anything.
pocket tts is completely polarizing. in the big benchmark thread with 180 timed runs it came out slowest of the cpu models tested but the latency stays flat no matter how long the text is and it streams. its fans are loud, they say nothing else comes close at its size for voice cloning and that it is like 90 percent as good as omnivoice at a fraction of the compute. it ranked first for cpu cloning against a bunch of tiny models. the haters just say it sounds terrible. honestly both camps are probably right, the quality bar is lower and the tradeoff is that it runs anywhere.
for what i run this on the verdict is fairly clean. the desktop has a big gpu, so omnivoice is the choice there, the non-commercial license does not matter for personal use and there is no quantized port yet anyway so you run the python server directly. if i ever need voice cloning on the weak cpu box instead, pocket tts is basically the only thing that clones voices at all and the flat latency thing is a genuinely cool trick for streaming. i know which one i am installing first.