No cloud, no tokens to pay. Reads a book using a full cast of characters. Quotation attribution detection (for my novel) is at 97.2% accuracy (485/499 quotes identified and assigned correctly). The autofill of character voice descriptions uses the prose to determine how the character sounds.
Employs Gemma 4[1] for the prose analysis (voice fills, quotation detection) and Qwen3 TTS Voice Design[2] for creating voice samples. Runs on an 8GB NVIDIA T1000 GPU card, 96 GB RAM, and a AMD Ryzen 5 7600.
I've been working on a similar project all year and as a tip, you should try Fish Audio or Higgs as a replacement for Qwen3. Both yield much better prosody and are much easier to listen to for long runs.
I wasn't able to find a version of these that can create voice samples based on voice designs. Do you mean to use Qwen3 TTS Voice Design to create samples followed by Higgs or Fish Audio to clone the sample voices and narrate the novel?
MOSS-TTS 2.0 will apparently have voice design, as well, on par with ElevenLabs quality.
For the voice design, these don’t support it, but for the final render, they’re much better. So your pipeline could for example generate voices with one tool and render with another.
It's cool technology and I read a lot of audiobooks, even hundreds of hours of TTS. I feel like my brain can fill in the character voices from the text - on the page it's not like they're different fonts.
I understand audiobook narrators often do it, and that's fun. But it's not so critical in my opinion
> annotate your text with extra 'stage directions'
Good idea, not something I've considered yet. Wouldn't take much to add it since there's already a feature for selecting a quotation and assigning it an intonation. Same infrastructure could be reused to select arbitrary text and assign stage directions.
FYI I solved that by changing the monitor - gigabyte m28u/m32u, built in KVM they run at 144 hertz. Only downside is that doesn't work easily if you want multi-monitor KVM.
Of possible interest, I once used an OSM editor to help draw Voronoi diagrams on spherical surfaces in a scalable vector graphic file format to help depict planet-wide fissures from colossal collisions.
"Studying these layered strata is known as magnetostratigraphy and can be used to verify radioactive dating methods. Geologists used magnetostratigraphy to cross-check the dates of the Nakali Formation, where Nakalipithecus nakayamai’s jaw was found."
27 KB is extraordinarily impressive. Here's a deterministic suite of common language- and format-specific PHP files an LLM wrote for me in 32 KB, for comparison:
The proof of finite-time singularity may impact both fluid dynamics models (CFD) and AI reasoning models. Under specific conditions, Navier–Stokes equations allow velocity to grow infinitely, causing the continuum fluid assumption to break down. Knowing the exact mathematical breakdown mechanisms helps developers improve adaptive mesh refinement and sub-grid scale models around high-vorticity regions (like vortex stretching and turbulent shear layers). While aerodynamic simulations for vehicles operate far from singularity thresholds, their stability at extreme boundaries could improve?
Proving out the combination of scaling inference-time compute and agent collaboration to solve previously intractable mathematical problems is WOW. By pairing creative candidate generation with automated proof checkers (like Lean) we are leaning into a repeatable framework for AI-driven scientific discovery.
> Knowing the exact mathematical breakdown mechanisms helps developers improve adaptive mesh refinement and sub-grid scale models around high-vorticity regions (like vortex stretching and turbulent shear layers).
This is 100% wrong and reads like copy paste of AI slop.
Any simulation which uses sub-grid scale models is already solving a different PDE than the actual Navier-Stokes considered in the Millenium problem, and that PDE is guaranteed to have different properties. Full stop.
And to claim this is somehow connected to AMR methods is an example of the kind of pseudoscientific statement Wolfgang Pauli would have called "not even wrong".
Why not just make it opt-in? Seems bad to leave the desires of nearly half your users unaddressed. As long as you're not shoehorning in features or making them unavoidable or the central focus, the only people who'd take issue would be zealot types who want to control other people's choices.
https://www.youtube.com/watch?v=WAeHgE94rVo
No cloud, no tokens to pay. Reads a book using a full cast of characters. Quotation attribution detection (for my novel) is at 97.2% accuracy (485/499 quotes identified and assigned correctly). The autofill of character voice descriptions uses the prose to determine how the character sounds.
Employs Gemma 4[1] for the prose analysis (voice fills, quotation detection) and Qwen3 TTS Voice Design[2] for creating voice samples. Runs on an 8GB NVIDIA T1000 GPU card, 96 GB RAM, and a AMD Ryzen 5 7600.
[1]: https://deepmind.google/models/gemma/gemma-4/
[2]: https://huggingface.co/spaces/Qwen/Qwen3-TTS-Voice-Design
reply