I've been using TTS systems for decades with accessibility use cases, so other than quality audio books that necessitate a skilled performing narrator, I no longer mind.
In fact, I prefer legacy Bayesian phonetic models over the newer convolutional and recurrent neural networks, as their hard consonants and robotic consistency in pronunciations and intonation are much easier to listen and discern at higher words per minute, like at 3x or 4x natural speech rates for everyday blind reading, as compared to modern mumbling/slurring of syllables or artificial stridor and other breathy sounds.
I haven't dug into Guix yet, so is the config more of a markup and less of Turing complete language? That sounds like it'd be easier to grock or optimize an LSP for.
I have heard that Guix takes a stronger stance with respect to unfree software. I don't think any of the official nix Hydra infrastructures build for unfree packages, but they are packaged and indexed into nixpkgs. Has Guix been difficult at all in that regard, i.e. using proprietary drivers or closed libraries for work or personal hardware?