Upload ai_ext.py

ai_ext.py

@@ -3,13 +3,14 @@ Imported by app_v2_entry.py to register /api/rewrite_share, /api/topic_post,
 /api/ai_wall, /api/wall, /api/ai/short endpoints on the main FastAPI app.
 
 Uses main.py's WALL_FILE (wall_posts.json) for unified data store.
+TTS: edge-tts (HoaiMy female, NamMinh male) with speed control + gTTS fallback.
 """
-import os, re, json, time, random, html as html_lib, subprocess
-from urllib.parse import quote_plus, quote, urlparse
-from typing import Optional
+import os, re, json, time, random, html as html_lib, subprocess, asyncio
+from urllib.parse import quote_plus, quote, urlparse, urljoin
+from typing import Optional, List, Dict
 import requests
 from bs4 import BeautifulSoup
-from fastapi import Request
+from fastapi import Request, Query
 from fastapi.responses import HTMLResponse, JSONResponse, FileResponse
 
 from main import app
@@ -18,7 +19,6 @@ from main import app
 try:
     from main import _load_wall, _save_wall, _web_context  # noqa: F401
 except ImportError:
-    # Fallback if main doesn't export them
     _data_dir = "/data" if os.path.isdir("/data") else "/app/data"
     _wall_file = os.path.join(_data_dir, "wall_posts.json")
     def _load_wall():
@@ -53,6 +53,10 @@ try:
     from PIL import Image, ImageDraw, ImageFont
 except Exception:
     Image = ImageDraw = ImageFont = None
+try:
+    import edge_tts
+except Exception:
+    edge_tts = None
 
 
 def _hf_token():
@@ -64,12 +68,48 @@ def _hf_token():
 
 HF_TOKEN = _hf_token()
 QWEN_VL_MODEL = os.getenv("QWEN_VL_MODEL", "Qwen/Qwen2.5-VL-7B-Instruct")
-SHORTS_DIR = os.path.join("/data" if os.path.isdir("/data") else "/app/data", "ai_shorts")
-HEADERS = {"User-Agent":"Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36","Accept-Language":"vi-VN,vi;q=0.9,en;q=0.8"}
+DATA_DIR = "/data" if os.path.isdir("/data") else "/app/data"
+SHORTS_DIR = os.path.join(DATA_DIR, "ai_shorts")
+HEADERS = {
+    "User-Agent": "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/125.0.0.0 Safari/537.36",
+    "Accept-Language": "vi-VN,vi;q=0.9,en;q=0.8"
+}
 LAST_QWEN_ERROR = ""
 
+# ===== TTS VOICE CONFIG =====
+TTS_VOICES = {
+    "hoaimy": {"id": "vi-VN-HoaiMyNeural", "gender": "female", "name": "Hoài My (Nữ)"},
+    "namminh": {"id": "vi-VN-NamMinhNeural", "gender": "male", "name": "Nam Minh (Nam)"},
+}
+TTS_DEFAULT_VOICE = "hoaimy"
+TTS_DEFAULT_SPEED = 1.2
+
+TOPIC_VOICE_MAP = {
+    "bóng đá": "namminh", "thể thao": "namminh", "world cup": "namminh",
+    "premier league": "namminh", "champions league": "namminh", "la liga": "namminh",
+    "serie a": "namminh", "bundesliga": "namminh", "v-league": "namminh",
+    "tennis": "namminh", "olympic": "namminh", "f1": "namminh", "moto": "namminh",
+    "sức khỏe": "hoaimy", "làm đẹp": "hoaimy", "giải trí": "hoaimy",
+    "âm nhạc": "hoaimy", "phim": "hoaimy", "thời trang": "hoaimy",
+    "ẩm thực": "hoaimy", "du lịch": "hoaimy", "gia đình": "hoaimy",
+    "tình yêu": "hoaimy", "hôn nhân": "hoaimy", "mẹ và bé": "hoaimy",
+    "công nghệ": "namminh", "ai": "namminh", "robot": "namminh",
+    "khoa học": "namminh", "vũ trụ": "namminh", "điện thoại": "namminh",
+    "laptop": "namminh", "game": "namminh",
+    "chính trị": "namminh", "kinh tế": "namminh", "tài chính": "namminh",
+    "chứng khoán": "namminh", "ngân hàng": "namminh", "thị trường": "namminh",
+    "xã hội": "namminh", "pháp luật": "namminh", "giáo dục": "namminh",
+}
+
+
+def _detect_voice_for_topic(title: str, text: str) -> str:
+    combined = (title + " " + text[:500]).lower()
+    for keyword, voice_id in TOPIC_VOICE_MAP.items():
+        if keyword in combined:
+            return voice_id
+    return TTS_DEFAULT_VOICE
+
 
-# ===== TEXT HELPERS =====
 def _clean_text(s: str) -> str:
     s = html_lib.unescape(s or "")
     return re.sub(r"\s+", " ", s).strip()
@@ -84,17 +124,13 @@ def _safe_name(s):
     return re.sub(r"[^a-zA-Z0-9_-]+", "_", str(s))[:80]
 
 
-# ===== CLEAN AI OUTPUT =====
 def _clean_ai_output(text: str) -> str:
-    """Remove markdown artifacts, instruction leakage, and deduplicate sentences."""
     if not text:
         return ""
-    # Remove markdown headings, bold, italic, horizontal rules
     text = re.sub(r'^#{1,6}\s+', '', text, flags=re.MULTILINE)
     text = re.sub(r'\*\*([^*]+)\*\*', r'\1', text)
     text = re.sub(r'\*([^*]+)\*', r'\1', text)
     text = re.sub(r'^---+\s*$', '', text, flags=re.MULTILINE)
-    # Remove common AI instruction leakage phrases
     leakage = [
         r'Dưới đây là', r'Theo yêu cầu', r'Tôi sẽ', r'Đây là',
         r'Nội dung', r'Nhiệm vụ', r'Vai trò', r'Tôi là',
@@ -102,7 +138,6 @@ def _clean_ai_output(text: str) -> str:
     ]
     for phrase in leakage:
         text = re.sub(r'^' + phrase + r'[^\n]*\n', '', text, flags=re.MULTILINE | re.IGNORECASE)
-    # Deduplicate consecutive identical lines
     lines = text.split('\n')
     result = []
     prev = ""
@@ -113,7 +148,6 @@ def _clean_ai_output(text: str) -> str:
         result.append(line)
         prev = stripped
     text = '\n'.join(result)
-    # Deduplicate sentences
     sentences = re.split(r'(?<=[.!?])\s+', text)
     seen_sents = set()
     unique_sents = []
@@ -127,6 +161,101 @@ def _clean_ai_output(text: str) -> str:
     return ' '.join(unique_sents).strip()
 
 
+# ===== EXTRACT ALL IMAGES FROM ARTICLE =====
+def _extract_all_images(soup, base_url: str) -> List[Dict]:
+    images = []
+    seen_urls = set()
+    skip_patterns = [
+        "avatar", "icon", "logo", "button", "banner-ad", "tracking",
+        "beacon", "pixel", "1x1", "spacer", "emoji", "sprite", "placeholder",
+        "advertisement", "ads", "widget", "sidebar", "footer-logo",
+    ]
+
+    def _add_image(src: str, alt: str = "", source_tag: str = "img"):
+        if not src or src.startswith("data:"):
+            return
+        abs_url = urljoin(base_url, src.strip())
+        if abs_url in seen_urls:
+            return
+        if any(p in abs_url.lower() for p in skip_patterns):
+            return
+        try:
+            parsed = urlparse(abs_url)
+            path = parsed.path.lower()
+            if any(path.endswith(ext) for ext in ['.svg', '.ico', '.gif']):
+                return
+        except Exception:
+            pass
+        seen_urls.add(abs_url)
+        images.append({"url": abs_url, "alt": alt, "source": source_tag})
+
+    # Strategy 1: Standard <img> tags with all lazy-load attributes
+    for img in soup.find_all("img"):
+        src = (img.get("src") or img.get("data-src") or img.get("data-lazy-src") or
+               img.get("data-original") or img.get("data-srcset", "").split(",")[0].strip().split(" ")[0])
+        _add_image(src, alt=img.get("alt", ""), source_tag="img")
+
+    # Strategy 2: srcset on <img>
+    for img in soup.find_all("img", srcset=True):
+        for part in img["srcset"].split(","):
+            part = part.strip()
+            if part:
+                _add_image(part.split(" ")[0], alt=img.get("alt", ""), source_tag="srcset")
+
+    # Strategy 3: <picture> with <source>
+    for picture in soup.find_all("picture"):
+        for source in picture.find_all("source"):
+            srcset = source.get("srcset", "")
+            for part in srcset.split(","):
+                part = part.strip()
+                if part:
+                    _add_image(part.split(" ")[0], source_tag="picture/srcset")
+        fallback_img = picture.find("img")
+        if fallback_img:
+            _add_image(
+                fallback_img.get("src") or fallback_img.get("data-src"),
+                alt=fallback_img.get("alt", ""),
+                source_tag="picture/img"
+            )
+
+    # Strategy 4: WordPress CMS patterns
+    for img in soup.find_all("img", class_=re.compile(r"wp-image|size-large|size-full|aligncenter")):
+        _add_image(img.get("data-src") or img.get("src"),
+                   alt=img.get("alt", ""), source_tag="wp-image")
+
+    # Strategy 5: Background images in style attributes
+    for tag in soup.find_all(style=re.compile(r"background-image")):
+        for m in re.findall(r'url\(["\']?(.*?)["\']?\)', tag.get("style", "")):
+            _add_image(m, source_tag="background-style")
+
+    # Strategy 6: og:image (featured/hero image)
+    og_image = soup.find("meta", property="og:image")
+    if og_image and og_image.get("content"):
+        _add_image(og_image["content"], source_tag="og:image")
+
+    # Strategy 7: twitter:image
+    tw_image = soup.find("meta", attrs={"name": "twitter:image"})
+    if tw_image and tw_image.get("content"):
+        _add_image(tw_image["content"], source_tag="twitter:image")
+
+    # Strategy 8: <figure> with <figcaption>
+    for figure in soup.find_all("figure"):
+        img = figure.find("img")
+        if img:
+            src = img.get("data-src") or img.get("src")
+            figcaption = figure.find("figcaption")
+            alt = figcaption.get_text(strip=True) if figcaption else img.get("alt", "")
+            _add_image(src, alt=alt, source_tag="figure")
+
+    # Strategy 9: <a> tags linking to images
+    for a in soup.find_all("a", href=True):
+        href = a["href"]
+        if any(href.lower().endswith(ext) for ext in [".jpg", ".jpeg", ".png", ".webp", ".gif"]):
+            _add_image(href, alt=a.get_text(strip=True)[:80], source_tag="link")
+
+    return images
+
+
 # ===== JINA READER =====
 def _reader_url(target_url: str) -> str:
     safe = quote(target_url, safe=":/?#[]@!$&'()*+,;=%")
@@ -134,7 +263,7 @@ def _reader_url(target_url: str) -> str:
 
 def jina_reader_markdown(url: str) -> str:
     jr = _reader_url(url)
-    r = requests.get(jr, headers={"Accept":"text/markdown,text/plain,*/*", "X-Return-Format":"markdown", "User-Agent":"Mozilla/5.0"}, timeout=35)
+    r = requests.get(jr, headers={"Accept": "text/markdown,text/plain,*/*", "X-Return-Format": "markdown", "User-Agent": "Mozilla/5.0"}, timeout=35)
     r.raise_for_status()
     return r.text or ""
 
@@ -165,10 +294,12 @@ def _parse_jina_markdown(md: str, url: str):
             break
     if not title and paras:
         title = paras[0][:90]
-    return {"url": url, "title": title or url, "summary": paras[0] if paras else "", "text": "\n".join(paras), "image": image, "images": [image] if image else [], "via": "jina"}
+    return {"url": url, "title": title or url, "summary": paras[0] if paras else "",
+            "text": "\n".join(paras), "image": image,
+            "images": [image] if image else [], "via": "jina"}
 
 
-# ===== WEB SCRAPE =====
+# ===== WEB SCRAPE (with full image extraction) =====
 def _best_content_block(soup):
     best, best_score = None, 0
     for el in soup.find_all(["article", "main", "section", "div"]):
@@ -200,6 +331,13 @@ def scrape_any_url_direct(url: str):
     image = img_tag.get("content", "") if img_tag else ""
     if image and image.startswith("//"):
         image = "https:" + image
+    all_images = _extract_all_images(soup, url)
+    image_urls = [img["url"] for img in all_images]
+    if image and image not in image_urls:
+        image_urls.insert(0, image)
+    elif image in image_urls:
+        image_urls.remove(image)
+        image_urls.insert(0, image)
     block = _best_content_block(soup) or soup
     paras, seen_p = [], set()
     for p in block.find_all("p"):
@@ -211,7 +349,11 @@ def scrape_any_url_direct(url: str):
             break
     if not title and paras:
         title = paras[0][:90]
-    return {"url": url, "title": title or url, "summary": paras[0] if paras else "", "text": "\n".join(paras), "image": image, "images": [image] if image else [], "via": _domain(url)}
+    return {
+        "url": url, "title": title or url, "summary": paras[0] if paras else "",
+        "text": "\n".join(paras), "image": image_urls[0] if image_urls else "",
+        "images": image_urls, "via": _domain(url)
+    }
 
 def scrape_any_url(url: str):
     """Try direct scrape first, fall back to Jina Reader."""
@@ -228,6 +370,8 @@ def scrape_any_url(url: str):
                     jr["title"] = data["title"]
                 if data.get("image"):
                     jr["image"] = data["image"]
+                if data.get("images"):
+                    jr["images"] = data["images"]
                 jr["via"] = data.get("via", _domain(url)) + " + jina"
                 return jr
     except Exception:
@@ -266,7 +410,7 @@ async def qwen_generate(prompt: str, image_url: Optional[str] = None, max_tokens
                 {"role": "system", "content": (
                     "Bạn là biên tập viên báo điện tử tiếng Việt. "
                     "Quy tắc cứng: "
-                    "(1) KHÔNG lặp lại câu/nội dung. "
+                    "(1) KHÔNG lặp lại câu/nội dung đã viết. Mỗi ý chỉ xuất hiện 1 lần. "
                     "(2) KHÔNG dùng Markdown (##, **, ---). "
                     "(3) KHÔNG viết 'Dưới đây là', 'Tôi sẽ', 'Theo yêu cầu', 'Nhiệm vụ', 'Vai trò'. "
                     "(4) KHÔNG bịa thông tin ngoài nguồn. "
@@ -287,6 +431,24 @@ async def qwen_generate(prompt: str, image_url: Optional[str] = None, max_tokens
     return None
 
 
+# ===== TTS GENERATION =====
+async def _generate_tts_edge(text: str, voice_id: str, speed: float, out_path: str):
+    """Generate TTS using edge-tts with voice and speed control."""
+    if edge_tts is None:
+        raise RuntimeError("edge-tts chưa cài đặt")
+    voice = TTS_VOICES.get(voice_id, TTS_VOICES[TTS_DEFAULT_VOICE])["id"]
+    pct = int((speed - 1.0) * 100)
+    rate = f"+{pct}%" if pct >= 0 else f"{pct}%"
+    communicate = edge_tts.Communicate(text, voice, rate=rate)
+    await communicate.save(out_path)
+
+def _generate_tts_gtts(text: str, out_path: str):
+    """Fallback TTS using gTTS (no voice/speed control)."""
+    if gTTS is None:
+        raise RuntimeError("gTTS chưa cài đặt")
+    gTTS(text, lang="vi").save(out_path)
+
+
 # ===== SHORT VIDEO GENERATION =====
 def _download_image(url, fallback_topic, out_path):
     if url:
@@ -342,7 +504,6 @@ def _make_short_frame(post, img_path, out_path):
     draw.rectangle((0, 780, W, H), fill=(14, 14, 14))
     draw.text((54, 830), "VNEWS · Tường AI", fill=(92, 184, 122), font=font_label)
     title = post.get("title", "")
-    # Wrap title
     words = title.split(); lines_t = []; cur = ""
     for w in words:
         if len(cur) + len(w) + 1 <= 24:
@@ -352,7 +513,6 @@ def _make_short_frame(post, img_path, out_path):
             cur = w
     if cur: lines_t.append(cur)
     draw.multiline_text((54, 900), "\n".join(lines_t[:3]), fill=(255, 255, 255), font=font_title, spacing=10)
-    # Wrap body
     body_text = post.get("text", "")
     words_b = body_text.split(); lines_b = []; cur_b = ""
     for w in words_b:
@@ -373,15 +533,12 @@ def _short_script(post):
         txt = txt[:800].rsplit(" ", 1)[0] + "."
     return f"{post.get('title', '')}. {txt}"
 
-def _generate_short_video(post, post_id: str) -> str:
-    """Generate MP4 short video. Returns video URL or empty string on failure."""
+async def _generate_short_video(post, post_id: str, voice_id: str = None, speed: float = None) -> str:
     try:
         os.makedirs(SHORTS_DIR, exist_ok=True)
         out_mp4 = os.path.join(SHORTS_DIR, _safe_name(post_id) + ".mp4")
         if os.path.exists(out_mp4):
             return "/api/ai/short-file/" + post_id
-        if gTTS is None:
-            return ""
         work = os.path.join(SHORTS_DIR, _safe_name(post_id))
         os.makedirs(work, exist_ok=True)
         img_path = os.path.join(work, "image.jpg")
@@ -390,7 +547,18 @@ def _generate_short_video(post, post_id: str) -> str:
         _download_image(post.get("img"), post.get("title", "AI news"), img_path)
         _make_short_frame(post, img_path, frame_path)
         script = _short_script(post)
-        gTTS(script, lang="vi").save(audio_path)
+        if voice_id is None:
+            voice_id = _detect_voice_for_topic(post.get("title", ""), post.get("text", ""))
+        if speed is None:
+            speed = TTS_DEFAULT_SPEED
+        try:
+            await _generate_tts_edge(script, voice_id, speed, audio_path)
+        except Exception as e:
+            print(f"[TTS edge-tts error] {e}, falling back to gTTS")
+            if gTTS:
+                _generate_tts_gtts(script, audio_path)
+            else:
+                return ""
         cmd = ["ffmpeg", "-y", "-loop", "1", "-i", frame_path, "-i", audio_path,
                "-shortest", "-c:v", "libx264", "-tune", "stillimage", "-pix_fmt", "yuv420p",
                "-c:a", "aac", "-b:a", "128k", "-vf", "scale=1080:1920", out_mp4]
@@ -402,25 +570,30 @@ def _generate_short_video(post, post_id: str) -> str:
 
 
 # ===== MAKE POST =====
-def make_post(title, text, image, source_url, kind, sources=None):
+def make_post(title, text, image, source_url, kind, sources=None, images=None):
     return {
         "id": str(int(time.time() * 1000)) + str(random.randint(100, 999)),
         "title": title, "text": text, "img": image, "url": source_url,
-        "kind": kind, "sources": sources or [], "video": "", "ts": int(time.time())
+        "kind": kind, "sources": sources or [], "video": "",
+        "images": images or [], "ts": int(time.time())
     }
 
 
 # ===== SHARED PROMPT BUILDER =====
-def _build_rewrite_prompt(title: str, raw: str) -> str:
+def _build_rewrite_prompt(title: str, raw: str, images: List[str] = None) -> str:
+    image_info = ""
+    if images and len(images) > 1:
+        image_info = f"\n\nBài viết có {len(images)} ảnh. Hãy tham khảo nội dung ảnh để bổ sung thông tin cho bài viết (nếu phù hợp)."
     return f"""Tóm tắt bài viết sau thành bài ngắn gọn đăng Tường AI (tối đa 200 từ).
 
-YÊU CẦU:
+YÊU CẦU QUAN TRỌNG:
 - Chỉ nêu các Ý CHÍNH, không viết chi tiết thừa.
-- KHÔNG lặp lại câu/nội dung.
+- KHÔNG lặp lại câu/nội dung. Mỗi ý chỉ xuất hiện 1 lần.
 - KHÔNG dùng Markdown (##, **, ---).
 - KHÔNG viết "Dưới đây là", "Tôi sẽ", "Theo yêu cầu".
-- Viết thành đoạn văn thuần, mạch lạc.
-- Giữ sự thật, không bịa.
+- Viết thành đoạn văn thuần, mạch lạc, dễ đọc.
+- Giữ sự thật, không bịa thông tin.
+- Tóm tắt súc tích, đúng trọng tâm.{image_info}
 
 Tiêu đề gốc: {title}
 
@@ -431,9 +604,9 @@ Nội dung gốc:
 def _build_topic_prompt(topic: str, ctx: str) -> str:
     return f"""Viết bài tổng hợp NGẮN GỌN về chủ đề: "{topic}" (tối đa 200 từ).
 
-YÊU CẦU:
+YÊU CẦU QUAN TRỌNG:
 - Chỉ tổng hợp các Ý CHÍNH từ nguồn.
-- KHÔNG lặp lại nội dung.
+- KHÔNG lặp lại nội dung. Mỗi ý chỉ xuất hiện 1 lần.
 - KHÔNG sao chép nguyên văn; diễn đạt lại.
 - KHÔNG dùng Markdown (##, **, ---).
 - KHÔNG viết "Dưới đây là", "Tôi sẽ", "Theo yêu cầu".
@@ -458,19 +631,18 @@ async def api_rewrite_share(request: Request):
     raw = (data.get("summary", "") + "\n" + data.get("text", "")).strip()
     if len(raw) < 60:
         return JSONResponse({"error": "Bài viết quá ngắn để tóm tắt"}, status_code=422)
-
-    prompt = _build_rewrite_prompt(data.get("title", ""), raw)
-    text = await qwen_generate(prompt, image_url=data.get("image") or None, max_tokens=700)
+    images = data.get("images", [])
+    prompt = _build_rewrite_prompt(data.get("title", ""), raw, images)
+    text = await qwen_generate(prompt, image_url=images[0] if images else None, max_tokens=700)
     if not text:
         return JSONResponse({"error": "Qwen2.5-VL chưa sẵn sàng: " + LAST_QWEN_ERROR}, status_code=503)
     text = _clean_ai_output(text)
-    post = make_post(data.get("title") or "Bài viết", text, data.get("image") or "", url, "rewrite")
-
-    # Auto-generate short video in background
-    video_url = _generate_short_video(post, post["id"])
+    post = make_post(data.get("title") or "Bài viết", text,
+                     images[0] if images else data.get("image", ""),
+                     url, "rewrite", images=images)
+    video_url = await _generate_short_video(post, post["id"])
     if video_url:
         post["video"] = video_url
-
     posts = _load_wall()
     posts.insert(0, post)
     _save_wall(posts)
@@ -490,18 +662,18 @@ async def api_url_wall(request: Request):
     raw = (data.get("summary", "") + "\n" + data.get("text", "")).strip()
     if len(raw) < 60:
         return JSONResponse({"error": "URL không có đủ nội dung"}, status_code=422)
-
-    prompt = _build_rewrite_prompt(data.get("title", ""), raw)
-    text = await qwen_generate(prompt, image_url=data.get("image") or None, max_tokens=700)
+    images = data.get("images", [])
+    prompt = _build_rewrite_prompt(data.get("title", ""), raw, images)
+    text = await qwen_generate(prompt, image_url=images[0] if images else None, max_tokens=700)
     if not text:
         return JSONResponse({"error": "Qwen2.5-VL chưa sẵn sàng: " + LAST_QWEN_ERROR}, status_code=503)
     text = _clean_ai_output(text)
-    post = make_post(data.get("title") or "Bài viết", text, data.get("image") or "", url, "url")
-
-    video_url = _generate_short_video(post, post["id"])
+    post = make_post(data.get("title") or "Bài viết", text,
+                     images[0] if images else data.get("image", ""),
+                     url, "url", images=images)
+    video_url = await _generate_short_video(post, post["id"])
     if video_url:
         post["video"] = video_url
-
     posts = _load_wall()
     posts.insert(0, post)
     _save_wall(posts)
@@ -514,11 +686,9 @@ async def api_topic_post(request: Request):
     topic = _clean_text(body.get("topic", ""))
     if not topic:
         return JSONResponse({"error": "missing topic"}, status_code=400)
-
     ctx = _web_context(topic)
     if not ctx:
         return JSONResponse({"error": "Không lấy được dữ liệu cho chủ đề này"}, status_code=422)
-
     image = pollinations_image_url(topic)
     prompt = _build_topic_prompt(topic, ctx)
     text = await qwen_generate(prompt, image_url=image, max_tokens=700)
@@ -526,11 +696,9 @@ async def api_topic_post(request: Request):
         return JSONResponse({"error": "Qwen2.5-VL chưa sẵn sàng: " + LAST_QWEN_ERROR}, status_code=503)
     text = _clean_ai_output(text)
     post = make_post(topic, text, image, "", "topic")
-
-    video_url = _generate_short_video(post, post["id"])
+    video_url = await _generate_short_video(post, post["id"])
     if video_url:
         post["video"] = video_url
-
     posts = _load_wall()
     posts.insert(0, post)
     _save_wall(posts)
@@ -540,37 +708,38 @@ async def api_topic_post(request: Request):
 # ===== WALL ENDPOINTS =====
 @app.get("/api/ai_wall")
 def api_ai_wall():
-    """Legacy endpoint — same data as /api/wall."""
     return JSONResponse({"posts": _load_wall()[:80]})
 
-
 @app.get("/api/wall")
 def api_wall():
-    """Unified wall endpoint — same data store as main.py's wall_posts.json."""
     return JSONResponse({"posts": _load_wall()[:80]})
 
 
+# ===== SHORT VIDEO ENDPOINT (with voice + speed params) =====
 @app.post("/api/ai/short/{post_id}")
-def api_ai_short(post_id: str):
-    """Generate (or retrieve cached) short video for a wall post."""
+async def api_ai_short(post_id: str, voice: str = Query(default=None), speed: float = Query(default=None)):
+    """Generate (or retrieve cached) short video for a wall post.
+    Query params:
+    - voice: 'hoaimy' (female) | 'namminh' (male) | auto-detect if not specified
+    - speed: float (default 1.2), e.g. 1.0=normal, 1.2=fast, 0.8=slow
+    """
     posts = _load_wall()
     post = next((p for p in posts if str(p.get("id")) == str(post_id)), None)
     if not post:
         return JSONResponse({"error": "post not found"}, status_code=404)
     os.makedirs(SHORTS_DIR, exist_ok=True)
     out_mp4 = os.path.join(SHORTS_DIR, _safe_name(post_id) + ".mp4")
-    if os.path.exists(out_mp4):
+    if os.path.exists(out_mp4) and voice is None and speed is None:
         video_url = "/api/ai/short-file/" + post_id
-        # Update post in store
         for i, p in enumerate(posts):
             if str(p.get("id")) == str(post_id):
                 posts[i]["video"] = video_url
                 break
         _save_wall(posts)
         return JSONResponse({"video": video_url})
-    if gTTS is None:
-        return JSONResponse({"error": "gTTS chưa sẵn sàng"}, status_code=503)
-    video_url = _generate_short_video(post, post_id)
+    if voice is not None and voice not in TTS_VOICES:
+        return JSONResponse({"error": f"voice không hợp lệ. Chọn: {list(TTS_VOICES.keys())}"}, status_code=400)
+    video_url = await _generate_short_video(post, post_id, voice_id=voice, speed=speed)
     if video_url:
         for i, p in enumerate(posts):
             if str(p.get("id")) == str(post_id):
@@ -596,5 +765,18 @@ def api_ai_status():
         "client_imported": AsyncInferenceClient is not None,
         "model": QWEN_VL_MODEL,
         "last_error": LAST_QWEN_ERROR,
-        "tts_ready": gTTS is not None
+        "tts_ready": gTTS is not None or edge_tts is not None,
+        "tts_engine": "edge-tts" if edge_tts else ("gtts" if gTTS else "none"),
+        "tts_voices": {k: v["name"] for k, v in TTS_VOICES.items()},
+        "tts_default_speed": TTS_DEFAULT_SPEED,
+    })
+
+
+@app.get("/api/ai/voices")
+def api_ai_voices():
+    return JSONResponse({
+        "voices": {k: {"name": v["name"], "gender": v["gender"]} for k, v in TTS_VOICES.items()},
+        "default_voice": TTS_DEFAULT_VOICE,
+        "default_speed": TTS_DEFAULT_SPEED,
+        "topic_voice_map": TOPIC_VOICE_MAP,
     })