Supertonic vs SuperClonica: same model, faster and cleaner
One model, Supertonic 3, with the same voices, the same text and the same seed. We voiced short chat phrases in Russian, English and German with stock Supertonic and with SuperClonica, then compared speed and sound.
In short: SuperClonica is 1.7 times faster on a graphics card and 1.6 times faster on a processor, starts speaking after 0.08 s instead of half a second, removes background noise in pauses and cuts off far fewer endings. Below are the measurements, the cases where stock is better, and every recording, so you can listen yourself.
Speed
Both engines ran in one process and took turns on each phrase, swapping the order every time, so background load on the computer slowed them down equally. Every call got a new seed, as in a chat where each phrase is new. Both engines were created in advance and warmed up.
On a graphics card the gain comes from CUDA graphs: the whole model is recorded once and then launched with a single command. On a processor the models are rewritten for multithreaded computation. SuperClonica was faster on all 10 phrases on the graphics card and on all 24 phrases on the processor.
| Supertonic | SuperClonica | |
|---|---|---|
| Graphics card, median per phrase | 202 ms | 117 ms |
| Graphics card, all 10 phrases in a row | 2.03 s | 1.19 s |
| Processor, median per phrase | 855 ms | 530 ms |
| Processor, 24 phrases in 8 languages in a row | 22.03 s | 14.39 s |
Sound
The model is the same, so the voice and intonation match. The difference is in what happens around the words: silence at the edges, noise in the pauses and the occasional stray sounds the model adds.
The chart shows one phrase on a shared time scale. Stock keeps half a second of silence before the first word and more than half a second after the last one; SuperClonica leaves 0.08 and 0.2 s, so speech starts sooner. Background noise in the pauses between words is removed too.
| On the 10 sample phrases | Supertonic | 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 |
| Total length of the 10 files | 37.6 s | 30.0 s |
| Background in pauses between words (median, relative to peak) | −48 dB | −92 dB |
What our tests on thousands of generations show
Ten phrases are a sample. These results come from much larger runs of our own checks.
- Stray sounds at the end of a phrase. After the last word the model sometimes adds a click, a breath, a hum or an extra syllable. In two runs of 1,800 and 1,392 phrases (12 and 18 languages, recognized with Whisper) there were 27 and 26 such endings. After SuperClonica's cleanup, 0 and 3 remained.
- Words are not cut. The cleanup is checked against the model's raw output with a phoneme recognizer. On recent versions that is more than 25,000 generations in 30+ languages with zero confirmed cuts of real speech.
- Fewer stumbles on long text. SuperClonica splits text into chunks of up to 160 characters at natural pauses; stock splits at 300. A repeated word happened once in 1,800 generations with chunks of up to 160 characters, and in 1–4% of cases with chunks of 200–290 characters.
- No clipping. The model sometimes produces peaks above full scale. A look-ahead limiter lowers only those peaks.
Fewer abrupt endings
The model sometimes cuts the last sound off at full volume, as if the recording were switched off mid-word. You hear it as a click or a swallowed ending. SuperClonica brings such endings to a natural fade-out and keeps the sound itself.
We voiced chat phrases with 4 standard voices each, with the same voice and seed for both engines. An abrupt ending is a drop in loudness of more than 25 dB within 2 ms at the end of speech. Below are three examples with the end of the phrase enlarged; the dashed line marks the end of speech.
| Abrupt cut at the end of a phrase | Supertonic | SuperClonica |
|---|---|---|
| Russian, 312 clips | 37 · 11.9% | 3 · 1.0% |
| English, 300 clips | 53 · 17.7% | 6 · 2.0% |
| German, 132 clips | 14 · 10.6% | 0 · 0.0% |
| Total, 744 clips | 104 · 14.0% | 9 · 1.2% |
Numbers
Stock passes digits to the model as they are. The model reads small numbers on its own but gets large ones wrong in some languages. SuperClonica spells numbers out as words in 27 languages.
Numbers were checked on a phrase with 3, 25, 2024 and 150, two seeds per language, recognized with Whisper.
| Language | Supertonic | SuperClonica |
|---|---|---|
| es | “150” turned into mush on both seeds | correct |
| fr | “150” sounds like “105” on both seeds | correct |
| de | on one seed “150” became “eine 50” | correct |
| pl | on one seed “2024” was garbled | correct |
| ru, uk, it | correct | correct |
Listen for yourself
All the phrases from the measurement. Under each recording is its loudness envelope, with the silence before and after speech shaded.
Where stock Supertonic is better
SuperClonica is not better at everything. These are the trade-offs we know about.
- Startup. On launch SuperClonica checks CUDA graphs on the graphics card and rewrites the models for the processor, so it starts slower: 9.5 s against 1.2 s on the graphics card and 4.0 s against 0.9 s on the processor. Create the engine once and keep it running.
- Long paragraphs on a graphics card. Text is split into chunks of up to 160 characters, so a long paragraph can come out slightly slower than stock; in our measurement this did not happen. In exchange there are fewer stumbles.
- Numbers without grammatical case. In Russian “в 2024 году” is read with the plain number instead of the ordinal form, and English years sound like ordinary numbers. The numbers=False parameter passes digits to the model as stock does.
How we measured
Supertonic 1.3.1 (the supertonic-3 model) without processing against SuperClonica 0.6.23 with default settings. Both engines ran at speed 1.05 with 8 steps, and each phrase had the same voice and seed in both. Graphics card: RTX 5070 Ti, ONNX Runtime GPU 1.26 (10 rounds). Processor: Ryzen 7 5800X, 4 threads for both, 24 chat phrases in 8 languages (2 rounds).
The computer was a working machine with applications open. On an idle machine the absolute milliseconds are lower, but the ratio holds. The sound metrics come from a separate run on the graphics card; silence and background are measured in 10 ms frames relative to the phrase's peak.
SuperClonica is an independent derivative of Supertonic, not an official Supertone release. The Supertonic 3 model is distributed under its own license.