Superclonica

Faster, cleaner speech from Supertonic 3: the same model and voices, with better sound and about 1.5x the speed of stock Supertonic on a GPU (1.6x on a CPU). 31 languages.

Custom voices for it - free and paid, or trained from your own recording - are at superclonica.com.

Install

For an NVIDIA GPU (Windows or Linux x86-64, Python 3.11+):

pip install superclonica-gpu-cuda

For any CPU (Windows, Linux, macOS, Python 3.10+):

pip install superclonica-cpu

Install one of them per environment: both provide the same superclonica module. To switch from the CPU package to the GPU one:

pip uninstall -y superclonica-cpu onnxruntime
pip install superclonica-gpu-cuda

Requirements

• GPU package: CUDA 12 and cuDNN come from pip; only an NVIDIA driver with CUDA 12 support is needed (R525 or newer; RTX 50-series cards need R570 or newer). Without a working GPU it falls back to the CPU with a warning.
• Linux: glibc 2.28 or newer (Ubuntu 20.04, Debian 10, RHEL 8 and later).
• macOS on Intel: Python 3.10-3.13; ONNX Runtime publishes no newer Intel macOS builds.
• Windows on ARM: Python 3.12 or newer; NumPy and SciPy publish no older ARM builds.

The model

The model (about 385 MB) is downloaded once on first use from its original Hugging Face repository, at a pinned revision, into ~/.cache/supertonic3 - the same folder the official supertonic package uses. Set SUPERTONIC_CACHE_DIR to keep it elsewhere, or pass an existing copy to work offline:

tts = Superclonica(model_dir="path/to/supertonic-3")

Quick start

from superclonica import Superclonica

tts = Superclonica()                     # GPU if available, else CPU
audio, sr = tts.synthesize("Hello! Nice to meet you.", voice="F3", lang="en", seed=7)
tts.save("hello.wav", audio)

From the command line:

superclonica --text "Good morning, everyone." --voice M1 --lang en --output hello.wav

Create the engine once and reuse it: start-up takes a few seconds.

Better than stock Supertonic

Sound

• Clean phrase edges. Small learned models remove clicks and noise at the start and end, and stray sounds the model sometimes adds after the last word (a click, a breath, an extra syllable, a hum) are cut away. A second pass over the finished phrase removes the faint breaths, thumps and clicks the model's hiss hid, and a piece that comes out empty or cut short is rendered again.
• Words stay whole. The cleanup keeps the words' own sounds - a final s or t that stands apart after its closure, a whispered first vowel, a last word after a pause - which is checked against the model's raw output with a phoneme recognizer on thousands of renders in 30+ languages.
• No clipping. The model often peaks above full scale; a look-ahead limiter lowers just those peaks.
• No dead air. The model pads 0.25-0.6 s of silence before speech and 0.6-0.9 s after it; this is trimmed to 80 ms and 200 ms, so speech starts sooner (trim_silence=False keeps the padding).
• Quiet background. A denoiser and a pause gate remove the model's hiss between words (about 40 dB lower than stock).
• Fewer stutters and repeated words. Long text is spoken in pieces of at most 160 characters with natural pauses; stock uses up to 300, where repeats become several times more frequent.

Text

• Numbers are spoken as words in 27 languages; stock misreads larger numbers in some languages.
• Any text works. Emoji, symbols and characters the model doesn't know are cleaned up instead of failing the request; words like "can't" and "well-known" stay intact.
• Abbreviations such as "pls", "brb", "idk" and "btw" are spelled out before speaking; stock reads them as made-up words. slang=False keeps them as written.

Speed

24 short phrases in 8 languages, same settings for both:

	Stock Supertonic	Superclonica
RTX 5070 Ti, median phrase	174 ms	117 ms
Ryzen 7 5800X (4 threads), median phrase	697 ms	432 ms

CUDA graphs on GPU, multi-threaded model rewrites on CPU, parallel lanes and batches.

Custom voices

The ten built-in voices are just the start. superclonica.com is a marketplace of custom Supertonic 3 voices, and they work with Superclonica as they are:

• try any voice online for free, and download the free ones without signing up;
• train a voice from your own 5-15 second recording, free, on your own NVIDIA GPU;
• or order a custom voice from one of the authors.

Each voice is a Supertonic voice-style JSON file:

audio, sr = tts.synthesize("Welcome back!", voice="my_voice.json", lang="en")

API

Superclonica(...)

Superclonica(model_dir=None, total_steps=None, speed=1.05, *,
             device="auto", device_id=0, threads=None, lanes=1,
             gpu_memory_mb=1024, cuda_graph="auto", chunking=True, numbers=True,
             denoise=True, target_lufs=-18.0, trim_silence=True, slang=True)

Parameter	What it does
model_dir	An existing Supertonic 3 copy (default: the automatic download above).
total_steps	Model steps per phrase (default 8).
speed	Speaking rate.
device	"auto" (GPU if it works, else CPU), "cuda" or "cpu".
device_id, gpu_memory_mb, cuda_graph	GPU number, memory limit per model session, CUDA graphs ("auto", True, False).
threads	CPU threads (default: one per physical core).
lanes	Independent model copies for the pieces of a long text and concurrent callers.
chunking, numbers, denoise, trim_silence, slang	Turn the text and sound steps above on or off.
target_lufs	Loudness used by normalize=True.

synthesize

synthesize(text, voice="M1", lang="en", *, seed=None, effect="auto", numbers=None,
           normalize=False, return_report=False)

Returns (audio, sample_rate): float32 mono at 44.1 kHz; with return_report=True also a dict describing the run. normalize=True brings the phrase to target_lufs; effect="off" disables the ending effects (the cleanup always runs).

synthesize_batch

Several texts at once, with the same options; voices, languages and seeds may be single values or one per text, and the results keep the input order:

results = tts.synthesize_batch(["First line.", "Second line."],
                               voices="F3", langs="en", seeds=[1, 2])
for audio, sr in results:
    print(len(audio) / sr, "seconds")

save

tts.save(path, audio) writes WAV (16-bit) or any other format soundfile supports.

Voices and languages

• Voices: M1-M5, F1-F5, or any Supertonic voice-style JSON, such as a custom voice from superclonica.com.
• Languages (31): ar bg cs da de el en es et fi fr hi hr hu id it ja ko lt lv nl pl pt ro ru sk sl sv tr uk vi

Command line

superclonica --text "Good morning!" --voice F3 --lang en --seed 7 --output morning.wav
superclonica --text-file story.txt --voice my_voice.json --lang en --output story.wav
superclonica --help

Responsible use

The Supertonic 3 model is licensed under the BigScience Open RAIL-M license. Its use restrictions (Attachment A in SUPERTONIC_MODEL_LICENSE.txt) apply to everyone using it through this package: among others, do not impersonate people without their consent, do not defame or harass, and disclose that generated content is machine-generated.

License

Apache-2.0, © 2026 superclonica.com. An independent derivative of Supertonic, not an official Supertone release. The bundled Supertonic SDK code is MIT; the Supertonic 3 model is under the BigScience Open RAIL-M license. See NOTICE.txt.

More about the project, custom voices and how to make your own: superclonica.com.