corrected Spanish VO

factory_config.yaml

@@ -47,7 +47,7 @@ pipeline:
   # Spanish: es-MX-JorgeNeural, es-MX-DaliaNeural, es-US-AlonsoNeural, es-ES-AlvaroNeural
   voices:
     en: "af_heart"
-    es: "es-MX-JorgeNeural"
+    es: "es-US-AlonsoNeural"
 
   # A/B Testing Thumbnails (generates N visual concepts in EN, ES, and Clean)
   thumbnail_count: 3

pipeline/script_gen.py

@@ -139,20 +139,29 @@ JSON Schema:
             "description": desc
         }
 
-        prompt = f"""You are a professional documentary translator.
+        prompt = f"""You are a master Spanish documentary voiceover narrator and localization expert (National Geographic / Discovery Channel en Español).
-Translate the following English YouTube documentary package into fluent, natural Latin-American Spanish for voiceover and YouTube upload.
+Translate the following English YouTube documentary package into captivating, natural, spoken Latin-American Spanish for voiceover narration and YouTube publishing.
 
 English Data:
 {json.dumps(translation_payload, ensure_ascii=False, indent=2)}
 
-Requirements:
+CRITICAL SPOKEN VOICE REQUIREMENTS:
-- "titles_es": 3 catchy Spanish titles matching the English titles
+1. Spoken Eloquence & Natural Flow:
-- "narrations_es": array of exact same length containing the Spanish voiceover text
+   - Write as if spoken by a native Latin-American documentary orator with rich, gripping cadence.
-- "chapters_es": array of exact same length containing short 2-4 word Spanish chapter titles
+   - Use natural spoken Spanish sentence connectors ("Sin embargo,", "A lo largo de los siglos,", "Lo más fascinante es que...", "No obstante,").
-- "on_screen_texts_es": array of exact same length containing Spanish translations of the on-screen text labels (keep empty if original is empty)
+   - Avoid rigid word-for-word literal translations from English.
-- "description_es": full Spanish translation of the YouTube description with Spanish hashtags
+2. Punctuation for Breathing & Timing:
-- Output STRICT JSON only with keys: {{"titles_es": [...], "narrations_es": [...], "chapters_es": [...], "on_screen_texts_es": [...], "description_es": "..."}}
+   - Use commas and question marks (¿...?) thoughtfully so the text-to-speech engine pauses with natural human breathing.
-- No markdown formatting or extra text."""
+3. Proper Nouns & Accuracy:
+   - Keep recognized proper names accurate ("Whitechapel", "Londres", "Alejandría", "Göbeklitepe").
+4. "titles_es": 3 high-CTR Spanish titles
+5. "narrations_es": array of exact same length containing the rich Spanish voiceover text
+6. "chapters_es": array of exact same length containing short 2-4 word Spanish chapter titles
+7. "on_screen_texts_es": array of exact same length containing Spanish lower-third labels
+8. "description_es": full Spanish description with Spanish hashtags
+
+STRICT OUTPUT FORMAT:
+Output STRICT JSON only: {{"titles_es": [...], "narrations_es": [...], "chapters_es": [...], "on_screen_texts_es": [...], "description_es": "..."}}"""
 
         for attempt in (1, 2):
             text = self._chat(pc, prompt, max_tokens=4096)
@@ -251,20 +260,25 @@ Return ONLY the corrected script as a single valid JSON object starting with {{
             ]
         }
 
-        prompt = f"""You are a master Spanish documentary editor and translator.
+        prompt = f"""You are a senior Spanish documentary voiceover director and phonetic editor.
-Review and polish this Spanish translation for topic "{topic}".
+Review, polish, and perfect this Spanish documentary narration for topic "{topic}".
 
-OBJECTIVES:
+OBJECTIVES FOR 100% NATURAL SPOKEN SPANISH:
-1. Fix proper nouns (keep standard names like "Whitechapel", "Jack el Destripador", "Londres", "Alejandría" accurate).
+1. Native Cadence & Phrasing:
-2. Correct any phonetic spelling mistakes or awkward phrasing.
+   - Eliminate any awkward literal translations or robotic sentence structures.
-3. Ensure natural Latin-American documentary narration tone and correct accent marks (tildes).
+   - Refine sentence flow so every line sounds natural, authoritative, and compelling when spoken aloud.
-4. Preserve the exact number of segments.
+2. Proper Punctuation & Breathing Pauses:
+   - Ensure natural comma placement to give the voice synthesizer natural breathing pauses.
+3. Proper Noun & Pronunciation Accuracy:
+   - Ensure historical figures, locations, and proper nouns are written in their correct standard Spanish forms (e.g., "Whitechapel", "Jack el Destripador", "Londres", "Alejandría", "Göbeklitepe").
+4. Preserve Structure:
+   - Maintain the exact same number of segments.
 
 Spanish Data:
 {json.dumps(payload, ensure_ascii=False, indent=2)}
 
 STRICT OUTPUT FORMAT:
-Return STRICT JSON only with keys: {{"titles_es": [...], "description_es": "...", "segments_es": [{{"chapter_title_es": "...", "text_es": "...", "on_screen_text_es": "..."}}, ...]}}"""
+Return STRICT JSON only: {{"titles_es": [...], "description_es": "...", "segments_es": [{{"chapter_title_es": "...", "text_es": "...", "on_screen_text_es": "..."}}, ...]}}"""
 
         try:
             raw_text = self._chat(pc, prompt, max_tokens=8192)

pipeline/tts_gen.py

@@ -4,6 +4,9 @@ Generates crystal-clear narration in English and Spanish with accurate per-segme
 """
 
 import os
+import sys
+import re
+import shutil
 import subprocess
 import logging
 import wave
@@ -38,8 +41,35 @@ try:
 except ImportError:
     edge_tts = None
 
-def _run_edge_tts(text: str, voice: str, out_wav_path: str):
+def clean_and_format_spanish_for_tts(text: str) -> str:
-    """Synthesize high-quality natural neural voiceover directly via module or CLI."""
+    """
+    Format Spanish narration text to sound 100% natural when spoken:
+    - Expands historical abbreviations (a.C., d.C., siglos, etc.)
+    - Formats numbers, units, and punctuation for natural human breathing pauses.
+    """
+    t = text.strip()
+    # Normalize historical dates and eras
+    t = re.sub(r'\b(?:a\.\s*C\.|a\s*C|A\.\s*C\.|A\s*C)\b', 'antes de Cristo', t)
+    t = re.sub(r'\b(?:d\.\s*C\.|d\s*C|D\.\s*C\.|D\s*C)\b', 'después de Cristo', t)
+    t = re.sub(r'\bsiglo\s+XXI\b', 'siglo veintiuno', t, flags=re.IGNORECASE)
+    t = re.sub(r'\bsiglo\s+XX\b', 'siglo veinte', t, flags=re.IGNORECASE)
+    t = re.sub(r'\bsiglo\s+XIX\b', 'siglo diecinueve', t, flags=re.IGNORECASE)
+    t = re.sub(r'\bsiglo\s+XVIII\b', 'siglo dieciocho', t, flags=re.IGNORECASE)
+    t = re.sub(r'\bsiglo\s+XV\b', 'siglo quince', t, flags=re.IGNORECASE)
+    t = re.sub(r'\bEE\.?\s*UU\.?\b', 'Estados Unidos', t)
+    t = re.sub(r'\bkm/h\b', 'kilómetros por hora', t)
+    t = re.sub(r'(\d+)\s*km\b', r'\1 kilómetros', t)
+    t = re.sub(r'(\d+)\s*m\b', r'\1 metros', t)
+    t = re.sub(r'(\d+)\s*ha\b', r'\1 hectáreas', t)
+    
+    # Soften ellipses and replace multiple hyphens with clean commas for natural pausing
+    t = re.sub(r'\.{3,}', ', ', t)
+    t = re.sub(r'\s*--+\s*', ', ', t)
+    t = re.sub(r'\s+', ' ', t)
+    return t
+
+def _run_edge_tts(text: str, voice: str, out_wav_path: str, rate: str = "-4%"):
+    """Synthesize high-quality natural neural voiceover directly via module or CLI with studio mastering."""
     global edge_tts
     if edge_tts is None:
         try:
@@ -51,28 +81,44 @@ def _run_edge_tts(text: str, voice: str, out_wav_path: str):
             except Exception as e:
                 logger.warning(f"[TTS] Dynamic edge-tts install failed: {e}")
 
+    # Preprocess Spanish text for maximum naturalness
+    if voice.startswith("es-") or "Neural" in voice:
+        spoken_text = clean_and_format_spanish_for_tts(text)
+    else:
+        spoken_text = text
+
     tmp_mp3 = out_wav_path.replace(".wav", "_temp.mp3")
     
     if edge_tts is not None:
         async def _speak():
-            communicate = edge_tts.Communicate(text, voice)
+            communicate = edge_tts.Communicate(spoken_text, voice, rate=rate)
             await communicate.save(tmp_mp3)
         asyncio.run(_speak())
     else:
         # Fallback to sys.executable -m edge_tts
-        cmd = [sys.executable, "-m", "edge_tts", "--voice", voice, "--text", text, "--write-media", tmp_mp3]
+        cmd = [sys.executable, "-m", "edge_tts", "--voice", voice, "--rate", rate, "--text", spoken_text, "--write-media", tmp_mp3]
         r = subprocess.run(cmd, capture_output=True, text=True)
         if r.returncode != 0:
             edge_bin = shutil.which("edge-tts") or os.path.expanduser("~/.local/bin/edge-tts") or os.path.expanduser("~/miniconda3/bin/edge-tts")
             if edge_bin and os.path.exists(edge_bin):
-                r2 = subprocess.run([edge_bin, "--voice", voice, "--text", text, "--write-media", tmp_mp3], capture_output=True, text=True)
+                r2 = subprocess.run([edge_bin, "--voice", voice, "--rate", rate, "--text", spoken_text, "--write-media", tmp_mp3], capture_output=True, text=True)
                 if r2.returncode != 0:
                     raise RuntimeError(f"edge-tts failed: {r.stderr or r2.stderr}")
             else:
                 raise RuntimeError(f"edge-tts execution error: {r.stderr}")
 
+    # Broadcast vocal mastering filter (warm low-mid EQ, subtle presence, transparent leveling)
+    vocal_af = (
+        "highpass=f=75,"
+        "equalizer=f=200:t=q:w=1.2:g=1.8,"
+        "equalizer=f=3500:t=q:w=1.5:g=1.0,"
+        "compand=attacks=0.03:decays=0.15:points=-80/-80|-40/-30|-20/-14|0/-2:gain=2,"
+        "volume=1.05"
+    )
+
     subprocess.run([
         "ffmpeg", "-y", "-i", tmp_mp3,
+        "-af", vocal_af,
         "-ar", "44100", "-ac", "2", "-c:a", "pcm_s16le", out_wav_path
     ], check=True, capture_output=True)
     if os.path.exists(tmp_mp3):
@@ -81,7 +127,7 @@ def _run_edge_tts(text: str, voice: str, out_wav_path: str):
 class TTSGen:
     def __init__(self, pcs: Dict[str, RemotePC], voices: Dict[str, str], tts_port: int = 8003):
         self.pcs = pcs
-        self.voices = voices or {"en": "af_heart", "es": "es-MX-JorgeNeural"}
+        self.voices = voices or {"en": "af_heart", "es": "es-US-AlonsoNeural"}
         self.tts_port = tts_port
 
     def generate(self, script_data: Dict[str, Any], job_dir: str, lang: str = "en") -> Dict[str, Any]: