SuperClonica — a faster, cleaner engine for Supertonic 3

SuperClonica runs the Supertonic 3 text-to-speech model with the same voices and the same seeds — and gets more out of it: 1.6–1.7× faster on CPU and GPU, speech that starts after 0.08 s instead of half a second, no background noise in pauses, far fewer cut-off endings and numbers spelled out correctly in 27 languages. It works offline on your own computer and reads any Supertonic 3 voice-style JSON file.

🗣️ Try it: online demo with free voices · ⬇️ free voices on GitHub · 🎙️ 170+ voices in the marketplace

Model	Supertonic 3 (about 99M parameters, ONNX), unchanged
Languages	31: en, ru, de, fr, es, pt, it, nl, pl, uk, cs, sk, sl, hr, bg, ro, hu, el, tr, ar, hi, id, vi, ja, ko, sv, da, fi, et, lv, lt
Voices	any Supertonic 3 voice-style JSON (style_ttl + style_dp), including Supertone's own M1–M5 and F1–F5
Runs on	CPU (Windows, macOS, Linux) or NVIDIA GPU (Windows, Linux x86-64)
Install	pip install superclonica-cpu or pip install superclonica-gpu-cuda, or the Docker image
Output	44.1 kHz mono WAV

Quick start

python -m pip install superclonica-cpu          # any CPU, Python 3.10+
# or, for an NVIDIA GPU (Python 3.11+; pip brings CUDA and cuDNN):
python -m pip install superclonica-gpu-cuda

Install only one of the two packages. The Supertonic 3 model (about 400 MB) downloads on the first run.

from superclonica import Superclonica

tts = Superclonica()  # GPU if available, otherwise CPU; create it once and keep it
audio, sample_rate = tts.synthesize(
    "Hello! Your new voice is ready.", voice="ciri.json", lang="en", seed=7
)
tts.save("hello.wav", audio)

Command line:

superclonica --text "Hello! Your new voice is ready." --voice ciri.json --lang en --output hello.wav

HTTP server in Docker (superclonica/superclonica):

docker run -d -p 8000:8000 -v superclonica-models:/models superclonica/superclonica:cpu
docker run -d --gpus all -p 8000:8000 -v superclonica-models:/models superclonica/superclonica:cuda

A free voice to start with: ciri.json and seven more in superclonica/supertonic-voices.

Speed

Same model, voice, text and seed for both engines, run in one process and taking turns on each phrase.

	Supertonic 3 (stock)	SuperClonica
GPU (RTX 5070 Ti), median per phrase	202 ms	117 ms
GPU, 10 phrases in a row	2.03 s	1.19 s
CPU (Ryzen 7 5800X, 4 threads), median per phrase	855 ms	530 ms
CPU, 24 phrases in 8 languages in a row	22.03 s	14.39 s

On the GPU the whole model runs as recorded CUDA graphs; on the CPU the models are rewritten for multithreaded computation. SuperClonica was faster on all 10 GPU phrases and all 24 CPU phrases.

Sound

The model is the same, so the voice and intonation match. The difference is around the words.

On 10 sample phrases	Supertonic 3	SuperClonica
Silence before the first word (median)	0.50 s	0.08 s
Silence after the last word (median)	0.63 s	0.20 s
Background in pauses (median, relative to peak)	−48 dB	−92 dB

Abrupt cut at the end of a phrase	Supertonic 3	SuperClonica
Russian, 312 clips	11.9%	1.0%
English, 300 clips	17.7%	2.0%
German, 132 clips	10.6%	0.0%

From larger runs of our own checks:

• Stray sounds after the last word (clicks, breaths, extra syllables): 27 and 26 in runs of 1,800 and 1,392 phrases with stock output; 0 and 3 after SuperClonica's cleanup.
• Words are never cut: the cleanup is checked against the raw output with a phoneme recognizer — more than 25,000 generations in 30+ languages with zero confirmed cuts of real speech.
• Fewer stumbles on long text: text is split at natural pauses into chunks of up to 160 characters (stock: 300).
• Numbers: spelled out as words in 27 languages, where the model alone misreads large numbers in some languages (numbers=False keeps stock behaviour).
• No clipping: a look-ahead limiter lowers only the peaks above full scale.

Every recording of the comparison, with its loudness envelope: Supertonic vs SuperClonica.

Where stock Supertonic 3 is better

• Startup: SuperClonica prepares CUDA graphs or CPU-optimized models on launch — 9.5 s vs 1.2 s on the GPU, 4.0 s vs 0.9 s on the CPU. Create the engine once and keep it running.
• Long paragraphs on a GPU can come out slightly slower because of the shorter chunks (not seen in our measurement).
• Numbers without grammatical case: Russian ordinals and English years are read as plain numbers; use numbers=False to pass digits to the model as stock does.

How it was measured

Supertonic 1.3.1 (supertonic-3) without processing against SuperClonica 0.6.23 with default settings, speed 1.05, 8 steps, the same voice and seed per phrase. GPU: RTX 5070 Ti, ONNX Runtime GPU 1.26 (10 rounds). CPU: Ryzen 7 5800X, 4 threads for both, 24 chat phrases in 8 languages (2 rounds). A working machine with applications open: absolute times are lower on an idle machine, the ratio holds.

Voices

Every Supertonic 3 voice is one JSON file of about 0.3 MB. Free ones are on GitHub and in the online demo; character, narrator and custom voices are in the SuperClonica marketplace, where every voice page lets you voice your own phrase online. You can also order a voice trained on your recording or train one for free on your own computer.

License

The SuperClonica engine is released under the Apache License 2.0 (pip show superclonica-cpu). It runs the Supertonic 3 model, which is distributed by Supertone under its own license (OpenRAIL-M); its terms apply to the model and to what you generate with it. SuperClonica is an independent derivative of Supertonic, not an official Supertone release.

Use voices responsibly: never to deceive, defraud or impersonate anyone (terms).

Links

Website · Install guide · PyPI: superclonica-cpu · PyPI: superclonica-gpu-cuda · Docker Hub · Python guide · Docker API guide