cosyvoice3 / Qwen3-TTS - specifying the target language during cloning

[subof]

Hi,
There is usually an option to select the target language when porting languages, but I don't find this option for these engines in Subtitle Edit.
At the same time, VoxCPM2 seems to be doing a good job without it, which is not the case with the other two.
Does it exist and how can I configure it?

Thanks

[CrispStrobe]

Yes — it's -tl <lang> on the CLI (or "language" in the POST /v1/audio/speech body); for cross-lingual cloning also pass -sl <lang> / "source_lang", which is the language your reference clip is spoken in.

You couldn't find it because it largely wasn't working: cosyvoice3 only switches to cross-lingual mode once it knows the reference's language, and that detector only understood Hangul/Kana/Han/Cyrillic — so every Latin-script pair (en→de, en→fr, es→it) silently stayed zero-shot and kept the reference's accent, while qwen3-tts discarded -tl outright.

Both are fixed on main (14be805). On real weights, an English reference asked for German now emits 122 speech tokens / 4.9 s instead of 262 / 10.5 s — the reference transcript was leaking into the output — and if the reference language still can't be determined, it now says so and points you at --source-lang rather than ignoring the request.

VoxCPM2 needs none of this because it's language-agnostic: it reads the script of the input text, so there's no knob to miss.

Full per-backend table of what the language actually does: docs/tts.md. Note Subtitle Edit will need to surface those two fields before you can set them from its UI.