Voice creation studio
Design a voice. Then direct it.
Resona creates voices from a description, a recording or a blend of two, keeps them in a library you search by meaning, and directs them through 75 emotions. It runs on your own CUDA GPU, on Qwen3-TTS.
The engine behind the audiobooks of Parlantia.
Pick an emotion and an intensity. Each emotion shifts three sampling parameters (temperature, repetition penalty, top-p); the intensity scales the shift from ×0.25 to ×2. These are the numbers the engine receives.
From a new voice to a finished audiobook.
Design
Describe a voice in words. The 1.7B VoiceDesign model renders it, and Resona keeps its reference audio and seed.
Clone
Start from a clean recording. The 0.6B Base model turns it into a 1024-dimension voice vector you can reuse.
Blend
Mix two voices with one weight and keep the lineage of the new one.
Library
Find a voice by what it sounds like, with local text embeddings on the CPU.
Dialogues
Give each line a voice and an emotion. Resona renders the lines as jobs, mixes them and masters the result: compression, then loudness to −14 LUFS.
Audiobooks and API
Assemble chapters into an .m4b with chapter marks. Drive it all from a REST API with asynchronous jobs.
Heard in Parlantia.
Parlantia publishes audiobooks of José Martí made with Resona. Three short excerpts, in Spanish. The full books are at parlantia.diaitu.com.
Audio: Parlantia, published there as samples.
Three processes, one GPU, your machine.
- An NVIDIA GPU with CUDA
- Python 3.12, uv and ffmpeg
- The Qwen3-TTS weights (1.7B VoiceDesign, 0.6B Base), downloaded once
Resona is in preview. To run it on your hardware, write to hello@syalia.com.
Preview. Resona is before its 1.0: interfaces still change and some modules are unfinished. The source will be public with a licence; this page will link it then.