The lead article: which open speech models actually cover English, Italian and Polish, which of them run well on GB10, and the constraint that forces a two-engine architecture.
Text-to-speech should be the easy workload on this machine. The models are small — half a billion parameters is typical, against the hundred billion the LLM series was wrestling with. They run comfortably faster than real time. You could hold a dozen of them in memory at once and never notice.
The difficulty is entirely linguistic. English is served by everything. Italian is served by most things. Polish is served by a noticeably smaller set, and the intersection of "clones voices well," "speaks Polish natively," and "has working support on ARM64 with CUDA 13" is small enough that the architecture is decided for you.
The best-supported TTS stack on this hardware is Qwen3-TTS. It is Apache 2.0, it comes in 0.6B and 1.7B variants, there are containers built specifically for GB10 with OpenAI-compatible APIs and CUDA-graph acceleration, and it runs at roughly 1.65–1.75× real time with first audio in about 0.4 seconds. It supports Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian.
It does not support Polish.
The best multilingual cloner that does cover all three of your languages is Chatterbox Multilingual v3 — MIT-licensed, 0.5B parameters, released 10 June 2026, covering 23+ languages including en, it and pl, cloning a voice from ten seconds of reference audio with no fine-tuning. It has no purpose-built GB10 packaging, and its dependency pins are actively hostile to this platform.
So you have a choice between one engine that is worse in English and Italian but covers everything, and two engines that are each strong where they are used. This series takes the second path — best-of-breed per language behind a single endpoint — because the extra complexity is about forty lines of routing code and the quality difference in daily use is not subtle.
| Model | Size / licence | en | it | pl | Cloning | GB10 support |
|---|---|---|---|---|---|---|
| Qwen3-TTS | 0.6B / 1.7B · Apache 2.0 | ✓ | ✓ | ✗ | 3–10 s reference, plus text-prompt "voice design" | Excellent — purpose-built containers, OpenAI API, CUDA graphs, streaming |
| Chatterbox Multilingual v3 | 0.5B · MIT | ✓ | ✓ | ✓ | 10 s reference, zero-shot, cross-lingual | Workable — needs manual ARM64 install; no official container |
| Chatterbox Turbo / Nano | 350M / 110M · MIT | ✓ | ✗ | ✗ | Yes | Good — Nano runs 3× real time on CPU alone |
| XTTS-v2 | ~0.5B · Coqui licence | ✓ | ✓ | ✓ | 6 s reference, the long-time reference point | Legacy — unmaintained since Coqui closed; check licence terms |
| Kokoro-82M | 82M · Apache 2.0 | ✓ | ✓ | ✗ | No — fixed voice set | Good, and very cheap — but no cloning and no Polish |
| F5-TTS | ~0.3B · MIT | ✓ | partial | via finetune | Yes, flow-matching, strong long-form prosody | Community builds; language coverage depends on the checkpoint |
| IndexTTS-2 | ~1B | ✓ | ✗ | ✗ | Yes, with duration and emotion control | Strong on English/Chinese; not a fit for this language mix |
Two things worth saying about this table. First, "supports language X" in a model card means the model was trained on that language — it does not promise the quality is even. Polish in a 23-language model is a smaller slice of the training data than English, and it sounds like it: expect occasional stress errors on longer words and a slight foreign-accent quality on some cloned voices. Second, licences here are unusually varied and unusually important, because voice is the one output type where the legal picture is genuinely active. Check them before anything leaves your house.
Recognition matters here even if you only want synthesis, for two reasons. It is how you verify your own output automatically (article 4 builds this loop: synthesise, transcribe, compare, flag), and it is how the video series gets timing information for lip sync and dubbing.
| Model | Character | For Polish |
|---|---|---|
| Whisper large-v3 | 1.55B, MIT, 99 languages, ~7.4% average WER on the open leaderboard | The safe default. Broadest language coverage available in the open, Polish well represented. |
| Parakeet TDT 0.6B v3 | NVIDIA, 25 European languages with automatic language ID, extraordinary throughput | Excellent fit: covers Polish and Italian, a quarter the size of Whisper, and dramatically faster. |
| Canary-Qwen 2.5B | Top of the open ASR leaderboard at ~5.63% average WER | Best raw accuracy, but leaderboard-leading results are English-centric — verify on your own Polish audio. |
| WhisperX | Whisper plus forced alignment and speaker diarization | What you want for subtitles and dubbing: word-level timestamps, not just text. There is a GB10 build. |
For this box the sensible pairing is Parakeet TDT v3 as the everyday recogniser — it is small, fast and covers both non-English languages — with WhisperX kept around for anything that needs word-level timing.
If "generative audio" extends past speech for you, ACE-Step 1.5 is the open music model that matters. It generates full compositions up to ten minutes, supports 50+ languages for sung lyrics, and runs in under 4 GB — trivially small for this machine. The XL series released in April 2026 added a 4B-parameter decoder for higher audio quality, in base, fine-tuned and turbo variants. It also supports LoRA personalisation from a handful of songs, cover generation, repainting and vocal-to-backing-track conversion.
It is also pre-bundled in the GB10 ComfyUI image the video series uses, which means if you build that, you get music generation without a separate installation. Convenient, and worth knowing before you set it up twice.
┌──────────────────────────────┐
your code ─────▶│ /v1/audio/speech (gateway) │
Open WebUI └──────────────┬───────────────┘
│ route on language
┌──────────────┴──────────────┐
▼ ▼
┌───────────────────┐ ┌────────────────────┐
│ Qwen3-TTS │ │ Chatterbox ML v3 │
│ en · it │ │ pl │
│ container, :8020 │ │ wrapped, :8030 │
└───────────────────┘ └────────────────────┘
│ │
└──────────────┬──────────────┘
▼
┌───────────────────────┐
│ Parakeet / WhisperX │
│ verification loop │
└───────────────────────┘One endpoint, one voice identity, three languages, everything local. Articles 2 and 3 build the two engines; article 4 builds the gateway, the quality loop and the UI.
Two things are worth being deliberate about, and one of them changes this week.
Consent. Voice is biometric-adjacent personal data. Cloning your own voice is unremarkable. Cloning someone else's — a colleague, a family member, a public figure — without their agreement is the kind of thing that is technically trivial, socially corrosive, and in several jurisdictions unlawful. The models make no distinction; you have to.
Labelling. Article 50 of the EU AI Act becomes applicable on 2 August 2026. It requires that AI-generated or AI-manipulated audio, video, image and text content be marked in a machine-readable way and that people be told when they are hearing synthetic audio, with a reinforced disclosure obligation for deepfakes — content resembling a real person that would plausibly be taken as authentic. The obligation applies whether or not there was intent to deceive. There is a narrow carve-out for evidently creative, satirical or fictional work. Penalties for transparency breaches run to €15 million or 3% of worldwide annual turnover.
For a private, on-box, personal-use setup this mostly does not bite — the obligations attach to providers and deployers putting systems and content into circulation, not to you talking to your own computer. It bites the moment anything you generate is published or sent to another person. The practical answer is to treat labelling as a property of your pipeline rather than an afterthought: