Supertonic 3 support

[jborza]

Supertonic 3 was released - https://github.com/supertone-inc/supertonic - with 31 supported languages. It would be awesome to integrate that, as I'd like to use it for more languages than English.

I see that in tts_supertonic.py we use the 'supertonic' module, which is supertonic 1.3.1, the new version 3 is used via onnxruntime.

[denizsafak]

I just tested them and totally agree with you, they sound awesome! I'm quite busy these days, so commits are welcome. Thanks for the heads up 😋

[jborza]

Was exploring it today, mostly for Czech and Slovak language and it's a generation ahead of previous voices, that's what got me interested. I vibe coded a synth app, seems to work, so i'll take a look at integrating (well, upgrading) into Abogen probably over the weekend.

[denizsafak]

That's awesome! Feel free to just change the code directly without asking, I trust you man. Let me know how the upgrade goes🙌

[Milor123]

+1000 haha awesome!! let me know, I want to try it too !! 😄

[jborza]

Supertonic has "voice" parameter and "language" parameter- I suppose we can add a dropdown to pick a language uf user picks Supertonic voice (e.g. M1)?

[Milor123]

Here I share with you some tests I did to convert books using supertonic.
The script detects the .txt files and then you pass it the argument of which one you want to convert.
In my opinion, it does a decent job in my language (Spanish).

convertir_libro_gpu.py (this is gpu version, its very fast, 26m for a long book [8 hours of speak])

convertir_libro.py (this is only CPU, this is slow 2h + for the same long book [8 hours of speak])

supertonic>=1.3.0
numpy>=1.21.0

for gpu in windows:

# ── GPU requirements for Supertonic 3 TTS ──
# Requires: NVIDIA GPU with CUDA 12.x, cuDNN 9, Windows
#
# Install:
#   pip install -r requirements_gpu.txt
#
# Then you MUST add these to PATH before running (the script handles this):
#   set PATH=%PATH%;.venv\Lib\site-packages\nvidia\cublas\bin
#   set PATH=%PATH%;.venv\Lib\site-packages\nvidia\cuda_nvrtc\bin
#   set PATH=%PATH%;.venv\Lib\site-packages\nvidia\cudnn\bin

supertonic>=1.3.0
onnxruntime-gpu>=1.23.1
nvidia-cudnn-cu12

I made the script quickly using Deepseek V4 flash.

[jborza]

yes, I also did some things mostly using code from supertonic sample - helper.py - although I did go for larger chunks than individual sentences.

There is also a total_steps parameter which helps on CPU only machines. Setting it to 5 or 6 is playable and faster on CPU machines than the default 8 - on my Ryzen 7 5800x the real-time factor for step count 5 = 0.13x, for 6 = 0.15x, for 8 = 0.21x, so it's 60% faster.

[jborza]

@Milor123 how were you able to get the GPU version running?
was it this?
import supertonic.config
supertonic.config.DEFAULT_ONNX_PROVIDERS[:] = [
"CUDAExecutionProvider",
"CPUExecutionProvider",
]

It requires cuDNN 9, I had CUDA 12.8 working for Abogen GPU version.

[jborza]

@Milor123 ooh, it's the pypi package, not the onnx version.

so yes, it works, but I had to set up manually - especially install onnxruntime-gpu

# ── 1. Poner DLLs de CUDA en el PATH ANTES de importar supertonic ──
venv_site = Path(__file__).parent / ".venv" / "Lib" / "site-packages"
nvidia_dirs = [
    venv_site / "nvidia" / "cublas" / "bin",
    venv_site / "nvidia" / "cuda_nvrtc" / "bin",
    venv_site / "nvidia" / "cudnn" / "bin",
    "c:\\Program Files\\NVIDIA\\CUDNN\\v9.22\\bin\\12.9\\x64" 
]

how did you get cudnn to get installed into venv?
(I see that abogen gpu installation does that with torch )

cheers

[Milor123]

Hi @jborza bro, check this:
supertonic_books_gpu_with_docs.zip Everything is documented here in English and I added the uv stuff
README_en.md (installation documentation in English)

[jborza]

@Milor123 it's super fast on GPU (4060) - synthesized 55 minutes of speech in 1:43 which means RTF of 0.031, which is 32x real time. I have to check Kokoro speed again.

[jborza]

@Milor123 out of curiosity, how good is the TTS in Spanish? for Slovak and Czech it's a lot better of existing free solutions (Piper).

[Milor123]

@Milor123 out of curiosity, how good is the TTS in Spanish? for Slovak and Czech it's a lot better of existing free solutions (Piper).
It is really good for standard terminology; I mean, for example, I use it for medical books and it works very well, but abbreviations like mg/dl and things like that have to be converted because it handles them very poorly. For instance, I have to change mg/dl to "milligrams per deciliter." Otherwise, I like it better than Kokoro in Spanish.

I have always used Balabolka with Microsoft SAPI; there is a female voice that does a very good job for medical books, but I love the speed that Supertonic has, I think I will adopt Supertonic.

[denizsafak]

It's pretty good with Turkish too. There is a Hugging Face Demo page and I think they all sound wonderful.

Also, I have a couple of ideas to clean up our tech stack and save us some development headaches:

1. Drop PyPI for ONNX: If we migrate our models to ONNX, we can shrink the overall app size and make adding new models easier. For example, the application could run these ONNX models in the background using WebGPU, similar to Supertonic's demo page. This might also eliminate our dependency on CUDA altogether.

2. Unify the UI: Right now, maintaining separate PyQt and web interfaces is a lot of double-work. We could switch to a hybrid setup like Tauri, NiceGUI, or Flask + PyWebView. That way, we write the UI once and it functions as both a web and desktop app.

Let me know your thoughts on this 🙌🏼

[jborza]

@denizsafak regarding PyPi and ONNX - that would mean migrating Kokoro as well. I'm not sure on WebGPU vs CUDA though, I think WebGPU would be much slower that CUDA on nvidia cards. That is not a big deal if you use TTS for articles, but it is if you convert books.

But I'd save that for other time, I believe progress is best made in steps and now the step (for this issue) would be to upgrade Supertonic to Supertonic 3.

Regarding 2. - absolutely.