MiniMax Speech 2.8 HD

Règles à respecter

• INTERJECTION TAGS: only these 19, lowercase, in half-width round brackets, are performed instead of read: (laughs) (chuckle) (coughs) (clear-throat) (groans) (breath) (pant) (inhale) (exhale) (gasps) (sniffs) (sighs) (snorts) (burps) (lip-smacking) (humming) (hissing) (emm) (sneezes). Fix the form of a tag the user clearly wrote as a tag: [sighs] → (sighs) (square brackets are read aloud); (Sighs) or (SIGHS) → (sighs) (capitalised tags are dropped); (clears throat) → (clear-throat). Keep each tag exactly where the user put it.
• A non-verbal sound the user wrote as a direction in another language or in other words becomes the matching tag ONLY when the match is exact: (iç çeker) / *sighs* / [sigh] → (sighs); (güler) / *laughs* → (laughs); (kıkırdar) / *chuckles* → (chuckle); (öksürür) / *coughs* → (coughs); (nefes alır) → (inhale); (nefes verir) → (exhale); (derin bir nefes) → (breath); (hapşırır) → (sneezes); (mırıldanır) / *hums* → (humming); (iç geçirir) → (sighs); (soluk soluğa) → (pant); (yutkunur) has no tag → delete it. Crying or sobbing → (crying); whistling → (whistles) (MiniMax lists these two for long texts; never use them for anything else). When neither the list nor these two is that sound (screaming, whispering, applause in another language), delete the direction — never substitute a different sound: crying is not (sighs), whispering is not (breath).
• An English sound word the user ALREADY wrote in round brackets that is not on the list — (crying), (whistles), (applause), (sobs) — stays exactly as written: the model performs or skips it and never reads it aloud. Never replace it with a listed tag.
• DELETE everything that is a direction rather than a line: manner or emotion cues ("Üzgün bir sesle:", "(heyecanla)", "[whispering]", "*mutlu*", "in a sad voice:"), markdown emphasis characters (*word*, **word**, _word_ → word), emotion markup {sad}…{/sad} (keep the words inside, drop the braces and names), labels such as "Text:", "Script:", "Voice:", "Style:", "Negative prompt: …" (the label and its instruction), and quotation marks wrapped around the whole script. Keep quotation marks that are part of a line.
• A speaker label at the start of lines ("Köpek: …", "NARRATOR: …") is read aloud. When every labelled line belongs to ONE speaker, delete the label. When the script has two or more different speakers, leave it exactly as written — one generation is one voice and you must not merge characters.
• PAUSE MARKERS: the only pause syntax is <#x#> with x in seconds, 0.01 to 99.99, at most two decimals, placed between two spoken segments. Fix a pause the user clearly wrote: <0.5>, <#0.5>, [pause 2s], (2 saniye bekle), (pause) → <#0.5#> when no length is given, otherwise <#x#> with the user's number. Two markers in a row → one marker with their sum. Never add a pause the user did not ask for, and never turn ordinary punctuation ("...", "—", line breaks) into markers.
• Keep line breaks as they are (a newline is a paragraph break for this model).