feat: Implement gender-specific TTS instructions, refactor async database session handling for character creation and preview generation, and add Aliyun voice design creation.

qwen3-tts-backend/core/audiobook_service.py

@@ -36,6 +36,26 @@ def _get_llm_service(user: User) -> LLMService:
     return LLMService(base_url=user.llm_base_url, api_key=api_key, model=user.llm_model)
 
 
+def _get_gendered_instruct(gender: Optional[str], base_instruct: str) -> str:
+    """Ensure the instruction sent to the TTS model has explicit gender cues if known."""
+    if not gender or gender == "未知":
+        return base_instruct
+    
+    # We want to force a clear gender bias at the start of the prompt
+    prefix = ""
+    if gender == "男":
+        prefix = "男性声音，"
+    elif gender == "女":
+        prefix = "女性声音，"
+    
+    if prefix and prefix not in base_instruct:
+        # Prepend prefix, but try to be smart if the first line starts with "音色信息："
+        if base_instruct.startswith("音色信息："):
+            return base_instruct.replace("音色信息：", f"音色信息：{prefix}", 1)
+        return f"{prefix}{base_instruct}"
+    return base_instruct
+
+
 def _extract_epub_chapters(file_path: str) -> list[str]:
     try:
         import ebooklib
@@ -166,6 +186,17 @@ async def analyze_project(project_id: int, user: User, db: Session, turbo: bool
 
         samples = _sample_full_text(text)
         n = len(samples)
+        
+        # Ensure previews directory is clean for new analysis
+        previews_dir = Path(settings.OUTPUT_DIR) / "audiobook" / str(project_id) / "previews"
+        if previews_dir.exists():
+            import shutil
+            try:
+                shutil.rmtree(previews_dir)
+            except Exception as e:
+                logger.warning(f"Failed to clear previews directory: {e}")
+        previews_dir.mkdir(parents=True, exist_ok=True)
+
         mode_label = "极速并发" if turbo else "顺序"
         ps.append_line(key, f"\n[LLM] 模型：{user.llm_model}，共 {n} 个采样段（{mode_label}模式），正在分析角色...\n")
         ps.append_line(key, "")
@@ -173,9 +204,12 @@ async def analyze_project(project_id: int, user: User, db: Session, turbo: bool
         def on_token(token: str) -> None:
             ps.append_token(key, token)
 
+        completed_count = 0
         def on_sample(i: int, total: int) -> None:
-            if i < total - 1:
-                ps.append_line(key, f"\n[LLM] 采样段 {i + 1}/{total} 完成，继续分析...\n")
+            nonlocal completed_count
+            completed_count += 1
+            if completed_count < total:
+                ps.append_line(key, f"\n[LLM] 采样段 {completed_count}/{total} 完成，继续分析...\n")
             else:
                 ps.append_line(key, f"\n[LLM] 全部 {total} 个采样段完成，正在合并角色列表...\n")
             ps.append_line(key, "")
@@ -202,33 +236,22 @@ async def analyze_project(project_id: int, user: User, db: Session, turbo: bool
 
         backend_type = user.user_preferences.get("default_backend", "aliyun") if user.user_preferences else "aliyun"
 
-        async def _create_char_with_voice(char_data):
+        for char_data in characters_data:
             name = char_data.get("name", "narrator")
             instruct = char_data.get("instruct", "")
             description = char_data.get("description", "")
             gender = char_data.get("gender") or ("未知" if name == "narrator" else None)
-
-            # Requires isolated DB queries since we're in an async concurrent block
-            try:
-                # We need an async wrapper or a local db session for concurrent sync DB pushes
-                # Because core crud uses synchronous SQLalchemy, executing them in threadpool via asyncio.to_thread
-                import asyncio
-                
-                def db_ops():
-                    from core.database import SessionLocal
-                    local_db = SessionLocal()
             try:
                 voice_design = crud.create_voice_design(
-                            db=local_db,
+                    db=db,
                     user_id=user.id,
                     name=f"[有声书] {project.title} - {name}",
                     instruct=instruct,
                     backend_type=backend_type,
                     preview_text=description[:100] if description else None,
                 )
-
                 crud.create_audiobook_character(
-                            db=local_db,
+                    db=db,
                     project_id=project_id,
                     name=name,
                     gender=gender,
@@ -236,18 +259,9 @@ async def analyze_project(project_id: int, user: User, db: Session, turbo: bool
                     instruct=instruct,
                     voice_design_id=voice_design.id,
                 )
-                    finally:
-                        local_db.close()
-                
-                await asyncio.to_thread(db_ops)
             except Exception as e:
                 logger.error(f"Failed to create char/voice for {name}: {e}")
 
-        import asyncio
-        batch_tasks = [_create_char_with_voice(cd) for cd in characters_data]
-        if batch_tasks:
-            await asyncio.gather(*batch_tasks)
-
         crud.update_audiobook_project_status(db, project_id, "characters_ready")
         ps.mark_done(key)
         logger.info(f"Project {project_id} character extraction complete: {len(characters_data)} characters")
@@ -259,25 +273,34 @@ async def analyze_project(project_id: int, user: User, db: Session, turbo: bool
         user_id = user.id
 
         async def _generate_all_previews():
-            async_db = SessionLocal()
+            # Get character IDs first using a temporary session
+            temp_db = SessionLocal()
             try:
-                db_user = crud.get_user_by_id(async_db, user_id)
-                characters = crud.list_audiobook_characters(async_db, project_id)
+                characters = crud.list_audiobook_characters(temp_db, project_id)
+                char_ids = [c.id for c in characters]
+            finally:
+                temp_db.close()
+            
+            if not char_ids:
+                return
 
             # Use a semaphore to limit concurrent TTS requests
             sem = asyncio.Semaphore(3)
+            
             async def _gen(char_id: int):
                 async with sem:
+                    # Each concurrent task MUST have its own dedicated session
+                    local_db = SessionLocal()
                     try:
-                            await generate_character_preview(project_id, char_id, db_user, async_db)
+                        db_user = crud.get_user_by_id(local_db, user_id)
+                        await generate_character_preview(project_id, char_id, db_user, local_db)
                     except Exception as e:
                         logger.error(f"Background preview generation failed for char {char_id}: {e}")
-                
-                tasks = [_gen(c.id) for c in characters]
-                if tasks:
-                    await asyncio.gather(*tasks)
                     finally:
-                async_db.close()
+                        local_db.close()
+            
+            tasks = [_gen(cid) for cid in char_ids]
+            await asyncio.gather(*tasks)
                 
         asyncio.create_task(_generate_all_previews())
 
@@ -538,7 +561,7 @@ async def generate_project(project_id: int, user: User, db: Session, chapter_ind
                         audio_bytes, _ = await backend.generate_voice_design({
                             "text": seg.text,
                             "language": "zh",
-                            "instruct": design.instruct,
+                            "instruct": _get_gendered_instruct(char.gender, design.instruct),
                         })
                 else:
                     if design.voice_cache_id:
@@ -563,7 +586,7 @@ async def generate_project(project_id: int, user: User, db: Session, chapter_ind
                             audio_bytes, _ = await backend.generate_voice_design({
                                 "text": seg.text,
                                 "language": "Auto",
-                                "instruct": design.instruct,
+                                "instruct": _get_gendered_instruct(char.gender, design.instruct),
                                 "max_new_tokens": 2048,
                                 "temperature": 0.3,
                                 "top_k": 10,
@@ -574,7 +597,7 @@ async def generate_project(project_id: int, user: User, db: Session, chapter_ind
                         audio_bytes, _ = await backend.generate_voice_design({
                             "text": seg.text,
                             "language": "Auto",
-                            "instruct": design.instruct,
+                            "instruct": _get_gendered_instruct(char.gender, design.instruct),
                             "max_new_tokens": 2048,
                             "temperature": 0.3,
                             "top_k": 10,
@@ -823,6 +846,20 @@ async def generate_character_preview(project_id: int, char_id: int, user: User,
             db.commit()
             logger.info(f"Bootstrapped local voice cache for preview: design_id={design.id}, cache_id={cache_id}")
 
+        if backend_type == "aliyun" and not design.aliyun_voice_id:
+            from core.tts_service import AliyunTTSBackend
+            if isinstance(backend, AliyunTTSBackend):
+                try:
+                    voice_id = await backend._create_voice_design(
+                        instruct=_get_gendered_instruct(char.gender, design.instruct),
+                        preview_text=preview_text,
+                    )
+                    design.aliyun_voice_id = voice_id
+                    db.commit()
+                    logger.info(f"Bootstrapped aliyun voice_id for preview: design_id={design.id}, voice_id={voice_id}")
+                except Exception as e:
+                    logger.warning(f"Failed to bootstrap aliyun voice_id for preview, falling back to instruct: {e}")
+
         if backend_type == "aliyun":
             if design.aliyun_voice_id:
                 audio_bytes, _ = await backend.generate_voice_design(
@@ -833,7 +870,7 @@ async def generate_character_preview(project_id: int, char_id: int, user: User,
                 audio_bytes, _ = await backend.generate_voice_design({
                     "text": preview_text,
                     "language": "zh",
-                    "instruct": design.instruct,
+                    "instruct": _get_gendered_instruct(char.gender, design.instruct),
                 })
         else:
             if design.voice_cache_id:
@@ -858,7 +895,7 @@ async def generate_character_preview(project_id: int, char_id: int, user: User,
                     audio_bytes, _ = await backend.generate_voice_design({
                         "text": preview_text,
                         "language": "Auto",
-                        "instruct": design.instruct,
+                        "instruct": _get_gendered_instruct(char.gender, design.instruct),
                         "max_new_tokens": 512,
                         "temperature": 0.3,
                         "top_k": 10,

qwen3-tts-backend/core/llm_service.py

@@ -122,12 +122,13 @@ class LLMService:
             "gender字段必须明确标注性别，只能取以下三个值之一：\"男\"、\"女\"、\"未知\"。\n"
             "narrator的gender固定为\"未知\"。\n"
             "对每个角色，instruct字段必须是详细的声音导演说明，需覆盖以下六个维度，每个维度单独一句，用换行分隔：\n"
-            "1. 音色信息：嗓音质感、音域、音量、气息特征（如：青年男性中低音，音色干净略带沙哑，音量偏小但稳定，情绪激动时呼吸明显）\n"
+            "1. 音色信息：嗓音质感、音域、音量、气息特征（例如，如果是女性角色，此处必须以'女性声音'开头，如：'女性声音，清脆悦耳的高音，嗓音纤细干净，带有一点点少女感'；男性角色则以'男性声音'开头）\n"
             "2. 身份背景：角色身份、职业、出身、所处时代背景对声音的影响\n"
             "3. 年龄设定：具体年龄段及其在声音上的体现\n"
             "4. 外貌特征：体型、面容、精神状态等可影响声音感知的特征\n"
             "5. 性格特质：核心性格、情绪模式、表达习惯\n"
             "6. 叙事风格：语速节奏、停顿习惯、语气色彩、整体叙述感\n\n"
+            "注意：instruct 的第一行（音色信息）必须与 gender 字段保持一致。如果 gender 为女，第一行绝对不能出现'男性'字样。\n\n"
             "只输出JSON，格式如下，不要有其他文字：\n"
             '{"characters": [{"name": "narrator", "gender": "未知", "description": "第三人称叙述者", "instruct": "音色信息：...\\n身份背景：...\\n年龄设定：...\\n外貌特征：...\\n性格特质：...\\n叙事风格：..."}, ...]}'
         )