# VoxCPM2 vs ZONOS3-Mini: Two Open Voice Models Tested in 24 Languages

> Source: <https://www.stork.ai/blog/voxcpm2-vs-zonos3-mini-open-tts-benchmark>
> Published: 2026-11-21 05:03:05+00:00

**Recommendation:** Score long speech clips in ten-second pieces at the pauses before trusting a recogniser's error rate; whole-clip scoring made complete Hindi and Urdu audio look 60% to 75% wrong for both models.
More in the playbook "Evals and testing for LLM apps": <https://wpnews.pro/playbooks/evals-testing>

[Make a voiceover](https://www.stork.ai/tools/text-to-speech?utm_source=house_ad&utm_medium=inline&utm_campaign=voiceover)

## The two models[#](#the-two-models)

| Sizes, language counts, licenses and price are from each maker's own page, read November 21, 2026. Memory and load time are our measurements. |  |  | 
|---|---|---|
|  | VoxCPM2 | ZONOS3-Mini | 
|---|---|---|
| Made by | OpenBMB | Zyphra | 
| Size | 2 billion parameters | 886 million parameters, about 100 million active per step | 
| Languages listed | 30 | 43 | 
| License | Apache-2.0 | Weights Apache-2.0, code MIT | 
| Voices | Invents a voice from a text description, or clones one from a clip | Clones a voice from a clip of up to 60 seconds; ships 27 ready voices | 
| GPU memory in our runs | 7.0 GB | 3.3 GB | 
| Time to load from disk in our runs | 76 seconds | 5.6 seconds | 
| Hosted by its maker | No | Yes, at $10.00 per 1 million UTF-8 bytes of text | 

VoxCPM2's details are from its [model card](https://huggingface.co/openbmb/VoxCPM2). ZONOS3-Mini's are from Zyphra's [model page](https://www.zyphra.com/models/zonos3-mini) and [price list](https://www.zyphra.com/open-intelligence); it was released on November 9, 2026, so it was less than two weeks old when we tested it. We ran the [transformers port](https://huggingface.co/Zyphra/ZONOS3-Mini-hf) at one pinned revision.

## Accuracy on an everyday script[#](#accuracy-on-an-everyday-script)

The first script in each language is the kind of thing people actually type: a 25-second birthday message from a grandmother, with a name, an age, a date, a question and an exclamation. Two voices, two takes each.

- **VoxCPM2 was more accurate in 14 of 16 languages** , and under 2.5% of characters wrong in 13 of them. In Spanish, Portuguese, French, Russian, Italian, Turkish, Polish and Indonesian it was under half a percent.
- **ZONOS3-Mini was ahead only in Urdu** (4.8% against 5.7%) and level in French. Its weakest results were Japanese (15.8%), Korean (7.0%), Hindi (6.5%) and Arabic (5.7%).
- **Score long clips in pieces.** Scored whole, the same Hindi and Urdu clips looked 60% to 75% wrong for both models, because the recogniser skipped the middle of them. Cut into ten-second pieces at the pauses, the audio turned out to be complete. Anyone benchmarking speech with a recogniser should check this before trusting a bad number.
- **Differences under one point are noise.** A recogniser makes its own mistakes, more of them outside the big European languages.

## Prices, numbers and abbreviations[#](#prices-numbers-and-abbreviations)

The second script is a product ad with a price, a count of 3,000 customers, a foreign brand name and an abbreviated title such as "Dr.". This is where the models differ most.

| Characters read back wrong on a product-ad script with a price, a number, a brand and an abbreviation. Two voices, two takes, scored in pieces. |  |  | 
|---|---|---|
| Language | VoxCPM2 | ZONOS3-Mini | 
|---|---|---|
| English | 2.0% | 0.9% | 
| Spanish | 7.3% | 6.2% | 
| Portuguese | 8.0% | 3.5% | 
| French | 2.1% | 2.8% | 
| Arabic | 6.3% | 36.2% | 
| Russian | 1.9% | 3.1% | 
| Hindi | 18.1% | 22.0% | 
| Italian | 3.3% | 7.0% | 
| Turkish | 1.8% | 6.4% | 
| German | 3.6% | 10.7% | 
| Urdu | 16.4% | 28.1% | 
| Japanese | 6.2% | 5.5% | 
| Korean | 6.8% | 9.7% | 
| Polish | 10.1% | 8.7% | 
| Indonesian | 2.4% | 4.7% | 
| Chinese | 1.7% | 1.2% | 

- **ZONOS3-Mini reads digits well only where it has a number grammar.** In English, Spanish, Portuguese, French, Italian, German, Japanese, Korean and Chinese it said the price correctly in every take. In Arabic, Russian, Hindi, Turkish and Indonesian it did not get the price right in any of four takes (in Hindi it sometimes read the digits out in English), and in Polish only in two. If you use it in those languages, write numbers out as words.
- **VoxCPM2 read the price correctly in 12 of 16 languages** in at least three of four takes. It failed on the Hindi price in every take and got the Urdu and Arabic prices right in only one take of four. On this script ZONOS3-Mini was ahead in six languages (English, Spanish, Portuguese, Japanese, Polish, Chinese) and VoxCPM2 in ten.
- **One Arabic take from ZONOS3-Mini stopped after 5.6 seconds** of a 23-second line. A length check with one retry, which any production pipeline should have, catches this.
- **Abbreviations trip both.** "Dott.ssa" in Italian and "MwSt." in German came out wrong more often than not. Spell them out.

## Very short lines[#](#very-short-lines)

The third script is three words ("Happy birthday, Mom!" in each language). Short lines are a known weak spot for voice models, and it shows:

- **VoxCPM2 read the three-word line cleanly in 14 of 16 languages.** The two misses were Hindi and Urdu.
- **ZONOS3-Mini got it wrong in at least one take in 8 of 16 languages** : English, Russian, Hindi, Italian, German, Urdu, Japanese and Chinese. The usual fault was added words after the line ended. In one German take it said the line twice and kept going; one English take ended "Happy birthday, Mom, and for".
- **The fix is the same for any model:** compare the length of the audio with what the text should take, and generate again when it runs long.

## How well each model keeps a voice[#](#how-well-each-model-keeps-a-voice)

Both models were given the same reference clip for each voice. We measured how close each output was to that clip with a speaker-recognition model (1.0 would be identical; two different speakers in our set scored between 0.04 and 0.36).

- **VoxCPM2 was closer to the reference in all 16 languages** , by 0.03 to 0.17.
- **Both stayed clearly the same speaker.** The lowest ZONOS3-Mini score, 0.71 in English, is still far above the different-speaker range.
- **This comparison favours VoxCPM2, and you should discount it accordingly.** The test voices were created with VoxCPM2's own voice design feature, and VoxCPM2 clones a clip together with its transcript, while we ran ZONOS3-Mini's lighter method (a speaker embedding). ZONOS3-Mini has a second cloning method we did not run.

Enjoying this? Get one like it in your inbox each morning.

one email a day · unsubscribe in two clicks · no third-party tracking

## Languages most open models skip[#](#languages-most-open-models-skip)

We also tested eight languages that VoxCPM2 does not list, because people type scripts in them every week: Persian, Ukrainian, Hungarian, Kazakh, Telugu, Nepali, Uzbek and Bengali. ZONOS3-Mini lists four of them (Ukrainian, Hungarian, Bengali, Telugu). One 10 to 15 second script per language, a woman's and a man's voice.

- **Hungarian and Kazakh are where ZONOS3-Mini changes the answer.** VoxCPM2 got about 14% of characters wrong in both; ZONOS3-Mini got 0% in Hungarian and 2.5% in Kazakh, which is not on its official list.
- **"Not listed" does not always mean "cannot".** VoxCPM2 read Persian and Telugu well, and ZONOS3-Mini read Persian, Kazakh and Nepali well, all unlisted.
- **Ukrainian:** both scored about 2%, but VoxCPM2 stuttered on the first word in both takes. On a longer test, ZONOS3-Mini read a two-minute Ukrainian script with 1.3% to 5% of characters wrong.
- **Bengali, Nepali and Uzbek cannot be judged by machine.** The recogniser gets about a fifth of real human speech wrong in these languages (the grey ticks), so the scores are only relative. On a longer Bengali script, ZONOS3-Mini's male voice read back 49% wrong. Do not ship these without a native listener.

### What else we tried for those eight languages[#](#what-else-we-tried-for-those-eight-languages)

| Other engines considered for the eight gap languages. Licenses read on each model's own page on November 20, 2026. |  |  | 
|---|---|---|
| Engine | License | What happened | 
|---|---|---|
| MOSS-TTS Local 1.7B | Apache-2.0 | Perfect on Hungarian; in Persian it repeated the reference clip's sentence instead of the script in 3 of 4 cloned takes. Used 13.3 GB of GPU memory. | 
| Kazakh and Uzbek adapters for VoxCPM2 | MIT | The Kazakh adapter fixed the woman's voice (2.1%) but not the man's (7.1%). The Uzbek one did not beat plain VoxCPM2. | 
| Piper voices (runs on a CPU) | Varies per voice: CC0, Apache-2.0, CC-BY-4.0, some CC-BY-SA | Accurate in Hungarian and Ukrainian, flat-sounding, and Persian has only male voices under a permissive license. | 
| OmniVoice | Weights CC-BY-NC | Not tested: non-commercial weights. | 
| Fish Audio S2 Pro | Research license | Not tested: no commercial use. | 
| XTTS-v2 | Coqui Public Model License | Not tested: non-commercial. | 
| Meta MMS-TTS | CC-BY-NC-4.0 | Not tested: non-commercial. | 

## Speed and what a minute of speech costs[#](#speed-and-what-a-minute-of-speech-costs)

| Self-hosting figures are our measurements on an NVIDIA A10 rented at $1.10 an hour. Hosted prices are our arithmetic on Zyphra's list price. |  |  | 
|---|---|---|
|  | VoxCPM2 | ZONOS3-Mini | 
|---|---|---|
| GPU seconds per second of speech (A10) | 0.71 to 0.83 | 0.37 to 0.51 | 
| GPU cost per minute of speech, model already loaded | about 1.4 cents | about 0.8 cents | 
| Load time before the first word | 76 s from disk (35 s from a memory snapshot) | 5.6 s from disk | 
| Maker's hosted price, English | not offered | about 0.8 cents a minute | 
| Maker's hosted price, Hindi | not offered | about 2 cents a minute | 

- **Zyphra charges by bytes of text, not by minutes of audio.** At[$10.00 per 1 million UTF-8 bytes](https://www.zyphra.com/open-intelligence) , English costs about 0.8 cents a minute at the speaking rate we measured. Hindi letters take three bytes each, so a minute of Hindi costs about 2 cents; Arabic and Russian, at two bytes a letter, land between 1 and 1.5 cents.
- **Self-hosting is only cheaper when the GPU stays busy.** The per-minute figures above ignore start-up and idle time. A single 15-second request that wakes a cold GPU and leaves it idle for 45 seconds afterwards costs two to three cents with VoxCPM2 on that GPU, most of it for waiting, not speaking.
- **ZONOS3-Mini's short load time matters more than its speed** if your traffic is bursty: six seconds to the first word against 35 to 76.

## What this test cannot tell you[#](#what-this-test-cannot-tell-you)

- **It did not measure how natural the voices sound.** A recogniser scores words, not accent, rhythm or warmth. A voice can read every word correctly and still sound foreign to a native speaker.
- **The voice comparison is tilted toward VoxCPM2** , for the reasons given above.
- **Two takes per cell is thin** , and a recogniser can fail on its own: our first scores for long Hindi and Urdu clips were wrong for that reason (see the correction at the top).
- **We ran one port of ZONOS3-Mini** , days after its release, with one of its two cloning methods and none of its emotion or pacing controls. Expect it to change.
- **The longest script was two minutes.** We did not test long narration.
- **One GPU type.** Speeds on other cards, and Zyphra's own optimised runtimes, will differ.

## Which one to use[#](#which-one-to-use)

- **One model for the major languages:** VoxCPM2. It was more accurate, steadier on short lines and numbers, and it can invent voices from a description.
- **Hungarian, Kazakh, Ukrainian and other languages outside VoxCPM2's list:** add ZONOS3-Mini. It is small enough to sit beside another model on one 24 GB GPU.
- **Cost-sensitive or bursty workloads in English and other Latin-alphabet languages:** ZONOS3-Mini, or Zyphra's hosted version, with a length check and a retry on short lines.
- **Hindi, Urdu, Bengali, Nepali:** test with native listeners whichever model you choose. Recognisers are least reliable in these languages, so machine scores alone prove little.
- **Whichever you pick:** write numbers and abbreviations out as words, check the audio length against the text, and regenerate when it is off.

We ran this benchmark while choosing voices for our own products. If you only need the result and not the GPUs, Stork's [text to speech](https://stork.ai/tools/text-to-speech) reads typed scripts in more than 30 languages, and [talking photo](https://stork.ai/tools/talking-photo) turns a photo and a script into a video.

## Frequently asked questions[#](#frequently-asked-questions)

### Is ZONOS3-Mini better than VoxCPM2?[#](#is-zonos3-mini-better-than-voxcpm2)

Not in general, in our tests. VoxCPM2 read everyday scripts more accurately in 14 of 16 major languages and kept the reference voice more faithfully in all 16. ZONOS3-Mini used a little over half the GPU time and was far better in languages VoxCPM2 does not list, such as Hungarian and Kazakh.

### Can I use VoxCPM2 and ZONOS3-Mini commercially?[#](#can-i-use-voxcpm2-and-zonos3-mini-commercially)

Both makers publish the weights under Apache-2.0, which allows commercial use; ZONOS3-Mini's code is MIT. Read the license on each model's own page before you ship, and check the weights' license separately from the code's for any other model you consider.

### How much does ZONOS3-Mini cost through Zyphra's API?[#](#how-much-does-zonos3-mini-cost-through-zyphras-api)

Zyphra lists it at $10.00 per 1 million UTF-8 bytes of input text (read November 21, 2026). At the speaking rates we measured that is about 0.8 cents per minute of English speech and about 2 cents per minute of Hindi, because Hindi letters take three bytes each.

### Which open text to speech model supports Hungarian, Kazakh or Ukrainian?[#](#which-open-text-to-speech-model-supports-hungarian-kazakh-or-ukrainian)

In our runs ZONOS3-Mini read Hungarian with 0% of characters wrong, Kazakh with 2.5% and Ukrainian with 1.3%, using a speech recogniser as the judge. Hungarian and Ukrainian are on its official list; Kazakh is not, but worked. VoxCPM2 got about 14% wrong in Hungarian and Kazakh.

### How much GPU memory do VoxCPM2 and ZONOS3-Mini need?[#](#how-much-gpu-memory-do-voxcpm2-and-zonos3-mini-need)

In our runs VoxCPM2 used 7.0 GB and ZONOS3-Mini 3.3 GB, both on an NVIDIA A10 with 24 GB. Both fit on one such GPU together.
