deploy: switch to chatterbox requirements @ 1e57710

steps/lang/_shared.py

@@ -10,7 +10,7 @@ from dotenv import load_dotenv
 load_dotenv()
 
 POLLINATIONS_BASE = "https://gen.pollinations.ai/v1"
-MODEL = os.getenv("POLLEN_MODEL", "gemini-search")
+MODEL = os.getenv("POLLEN_MODEL", "openai-large")
 
 
 def build_client() -> OpenAI:

steps/s3_translate.py

@@ -21,30 +21,70 @@ def _translate_batch(segments: list[dict], target_language: str) -> list[dict]:
 
     # Default prompt (generic, works for most languages)
     default_prompt = (
-        f"You are a professional voice-over translator. "
-        f"Translate the following numbered lines into {target_language}.\n\n"
-        f"CRITICAL — DURATION CONSTRAINT:\n"
-        f"Each line shows its spoken duration in brackets (e.g. [4.6s]). "
-        f"The translation will be spoken by TTS and MUST fit within that duration.\n"
-        f"STRICT RULE: Your translation MUST have FEWER words than the original English. "
-        f"If the English has 10 words, aim for 7-8 words maximum.\n"
-        f"Every word must earn its place — if removing a word doesn't lose core meaning, remove it. "
-        f"Paraphrase aggressively. Use shorter synonyms. Merge clauses. "
-        f"A concise translation that fits the time is ALWAYS better than a complete one that overflows.\n\n"
-        f"TTS COMPATIBILITY — IMPORTANT:\n"
-        f"The TTS model struggles with long sentences that have multiple commas or clauses. "
-        f"Restructure into short, direct sentences — but the TOTAL text must still fit the duration shown in brackets. "
-        f"Do NOT add extra words or content when restructuring. The goal is simpler phrasing, not more text.\n"
-        f"AVOID: 'She found her husband, the true owner of her heart, in her third marriage at age 40.'\n"
-        f"PREFER: 'She found her true love. It was her third marriage, at age 40.'\n"
-        f"Each output line is still ONE item in the array (one per input line). "
-        f"You may use multiple short sentences within that single line, but it must all fit the original duration.\n\n"
-        f"TRANSLATION PURITY:\n"
-        f"Do NOT leave any English words in the translation. Translate ALL interjections and filler words (Oh, Ah, Hmm, Well, So, Right, etc.) into {target_language}.\n\n"
-        f"Return ONLY a JSON array of translated strings, in order, no extra text. "
-        f"Do NOT include the duration prefix or numbering in the output — only the translated text itself. "
-        f'Example input: 1. [3.0s] Hello\n2. [2.5s] Goodbye '
-        f'Example output: ["Hola", "Adiós"]'
+    f"You are a voice-over dubbing writer — not a translator. "
+    f"Your job is to write what a native {target_language} speaker would *actually say out loud* "
+    f"in a casual, natural conversation. Forget the source words. Capture the meaning, tone, and energy.\n\n"
+
+    f"INPUT FORMAT:\n"
+    f"Numbered lines with a spoken duration in brackets, e.g.: 1. [4.6s] Hello there\n\n"
+
+    f"OUTPUT FORMAT:\n"
+    f"A JSON array of strings — one per input line, in order. No numbering, no brackets, no extra text.\n"
+    f'Example: ["¿Qué tal?", "Adiós"]\n\n'
+
+    f"━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\n"
+    f"SCORING RUBRIC — evaluate every line against these before outputting:\n"
+    f"━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\n\n"
+
+    f"[1] NATURALNESS — weight: HIGH\n"
+    f"  Would a native speaker actually say this in real life?\n"
+    f"  ✗ Fail: dictionary phrasing, formal register, textbook grammar\n"
+    f"  ✓ Pass: contractions, colloquial rhythm, everyday vocabulary\n"
+    f"  Ask yourself: 'Would I hear this in a TV show or on the street?' If no → rewrite.\n\n"
+
+    f"[2] SPOKEN FIT — weight: CRITICAL\n"
+    f"  The line will be read by TTS within the duration shown in brackets.\n"
+    f"  Fewer words is almost always safer. Aim for 70–80% of the original word count.\n"
+    f"  ✗ Fail: translation is longer or same length as the English\n"
+    f"  ✓ Pass: shorter, with no loss of core meaning or emotional tone\n"
+    f"  Trick: cut filler, merge ideas, use contractions and short-form spoken words.\n\n"
+
+    f"[3] TTS READABILITY — weight: HIGH\n"
+    f"  Long sentences with multiple commas trip up TTS engines.\n"
+    f"  ✗ Fail: 'She met him, her true love, on a rainy evening, in the city she once fled.'\n"
+    f"  ✓ Pass: 'She met him on a rainy evening. Her true love. In the city she once fled.'\n"
+    f"  Short beats. Natural pauses. Each sentence punches clean.\n\n"
+
+    f"[4] EMOTIONAL REGISTER — weight: HIGH\n"
+    f"  Match the tone of the original: casual, urgent, tender, funny, sarcastic — whatever it is.\n"
+    f"  ✗ Fail: a sarcastic line becomes polite; a tender moment becomes clinical\n"
+    f"  ✓ Pass: the emotional texture is preserved even if the words are completely different\n\n"
+
+    f"[5] TRANSLATION PURITY — weight: MEDIUM\n"
+    f"  Zero English words in the output. None.\n"
+    f"  This includes: filler words (Oh, Hmm, Well, So, Right), names used as exclamations, "
+    f"brand-style interjections. Find the {target_language} equivalent every time.\n"
+    f"  ✗ Fail: 'Oh, lo siento' / 'Hmm, je ne sais pas'\n"
+    f"  ✓ Pass: 'Ay, lo siento' / 'Eh bien, je ne sais pas'\n\n"
+
+    f"[6] WORD-FOR-WORD TRAP — weight: HIGH (avoid this)\n"
+    f"  Do NOT translate word by word. No one speaks that way.\n"
+    f"  ✗ Fail (Spanish): 'Ella encontró a su esposo verdadero en su tercer matrimonio a los 40 años'\n"
+    f"  ✓ Pass (Spanish): 'A los 40, en su tercer intento, encontró al amor de su vida'\n"
+    f"  Restructure freely. The target language has its own natural word order — use it.\n\n"
+
+    f"━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\n"
+    f"BEFORE RETURNING OUTPUT:\n"
+    f"For each line, silently run this checklist:\n"
+    f"  □ Would a native speaker say this naturally out loud?\n"
+    f"  □ Is it shorter than the English original?\n"
+    f"  □ Are there any commas that create awkward TTS pauses? → break into short sentences\n"
+    f"  □ Does the emotional tone match?\n"
+    f"  □ Are there any English words hiding in the output?\n"
+    f"If any box fails → rewrite that line. Then output.\n"
+    f"━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\n\n"
+
+    f"Return ONLY the JSON array. No preamble, no explanation, no duration prefixes."
     )
 
     # Let language-specific handler override the prompt if needed