fix(ai-avatar): 端到端一致性修复（标题/封面/B-roll位置/B-roll时长） (#1869)

Co-authored-by: frontend-dev <frontend-dev@xiaoxiajianji.com>
Co-committed-by: frontend-dev <frontend-dev@xiaoxiajianji.com>

alembic/versions/075_add_sentence_timings.py

@@ -0,0 +1,27 @@
+"""add sentence_timings to lipsync_jobs
+
+Revision ID: 075_add_sentence_timings
+Revises: 074_ai_avatar_render_script_id_optional
+Create Date: 2026-09-12
+"""
+
+import sqlalchemy as sa
+
+from alembic import op
+
+revision = "075_add_sentence_timings"
+down_revision = "074_render_script_id_optional"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    with op.batch_alter_table("lipsync_jobs") as batch:
+        batch.add_column(
+            sa.Column("sentence_timings", sa.JSON(), nullable=True),
+        )
+
+
+def downgrade() -> None:
+    with op.batch_alter_table("lipsync_jobs") as batch:
+        batch.drop_column("sentence_timings")

apps/api/app/schemas/lipsync.py

@@ -33,6 +33,7 @@ class LipsyncJobResponse(BaseModel):
     output_duration: float
     error_message: str
     error_code: str
+    sentence_timings: Optional[list] = None
     submitted_at: Optional[datetime] = None
     completed_at: Optional[datetime] = None
     created_at: datetime

apps/api/app/services/ai_avatar_render_service.py

@@ -25,6 +25,7 @@ from packages.adapters.sqlalchemy_impl.models import (
     ScriptModel,
 )
 from packages.domain.video_filter_builder import (
+    build_broll_overlay_filter,
     build_title_drawtext_filter,
 )
 from packages.shared.storage import get_shared_storage_service
@@ -228,27 +229,49 @@ class AiAvatarRenderService:
             self.db.commit()
 
             # 2. 构建 FFmpeg 滤镜链 (40%)
-            from packages.domain.video_filter_builder import build_broll_overlay_filter
+            # 用 ffprobe 探测输入视频分辨率，确保 B-roll 缩放与标题位置与实际输出一致。
+            # AI 数字人对口型输出为 9:16 竖屏，默认兜底 720x1280；探测失败时使用默认值不阻断渲染。
+            output_width, output_height = self._probe_video_resolution(input_video_path)
+            if output_width <= 0 or output_height <= 0:
+                output_width, output_height = 720, 1280
+                logger.info(
+                    "[数字人渲染] ffprobe 探测分辨率失败或无效，使用默认竖屏尺寸 %sx%s",
+                    output_width,
+                    output_height,
+                )
+            else:
+                logger.info("[数字人渲染] 探测输入视频分辨率: %sx%s", output_width, output_height)
 
-            filter_complex = build_broll_overlay_filter(
+            broll_filter, broll_label = build_broll_overlay_filter(
                 b_roll_segments=job.b_roll_segments,
                 video_duration=lipsync_job.output_duration,
+                output_width=output_width,
+                output_height=output_height,
             )
 
-            # 标题叠加
-            title_filter = build_title_drawtext_filter(job.title_config)
-            if title_filter:
-                if filter_complex:
-                    filter_complex += f"[vout]{title_filter}[vout_titled];"
-                else:
-                    filter_complex = f"[0:v]{title_filter}[vout_titled];"
+            # 标题叠加（传入实际输出尺寸，保证位置计算正确）
+            title_filter = build_title_drawtext_filter(
+                job.title_config,
+                output_width=output_width,
+                output_height=output_height,
+            )
 
-            # 清理末尾分号
-            if filter_complex.endswith(";"):
-                filter_complex = filter_complex[:-1]
-
-            # 最终输出标签
-            final_label = "vout_titled" if title_filter else ("vout" if filter_complex else None)
+            filter_complex = ""
+            final_label = None
+            if broll_filter and title_filter:
+                # B-roll → 标题叠在 B-roll 输出上
+                filter_complex = broll_filter + f";[{broll_label}]{title_filter}[vout_titled]"
+                final_label = "vout_titled"
+            elif broll_filter:
+                filter_complex = broll_filter
+                final_label = broll_label
+            elif title_filter:
+                filter_complex = f"[0:v]{title_filter}[vout_titled]"
+                final_label = "vout_titled"
+            else:
+                # 无滤镜：直接拷贝视频流
+                filter_complex = ""
+                final_label = None
 
             job.progress = 40
             self.db.commit()
@@ -427,6 +450,37 @@ class AiAvatarRenderService:
                 os.unlink(tmp.name)
             raise
 
+    @staticmethod
+    def _probe_video_resolution(video_path: str) -> tuple[int, int]:
+        """用 ffprobe 探测视频分辨率，返回 (width, height)；失败返回 (0, 0)。"""
+        try:
+            result = subprocess.run(
+                [
+                    "ffprobe",
+                    "-v",
+                    "error",
+                    "-select_streams",
+                    "v:0",
+                    "-show_entries",
+                    "stream=width,height",
+                    "-of",
+                    "csv=p=0:s=x",
+                    video_path,
+                ],
+                capture_output=True,
+                text=True,
+                timeout=15,
+            )
+            if result.returncode == 0 and result.stdout.strip():
+                parts = result.stdout.strip().split("x")
+                if len(parts) == 2:
+                    w, h = int(parts[0]), int(parts[1])
+                    if w > 0 and h > 0:
+                        return w, h
+        except Exception as exc:
+            logger.warning("[数字人渲染] ffprobe 探测分辨率失败: %s", exc)
+        return 0, 0
+
     def _build_ffmpeg_command(
         self,
         *,
@@ -450,7 +504,16 @@ class AiAvatarRenderService:
                 cmd.extend(["-i", asset_url])
 
         if filter_complex and final_label:
-            cmd.extend(["-filter_complex", filter_complex, "-map", f"[{final_label}]"])
+            cmd.extend(
+                [
+                    "-filter_complex",
+                    filter_complex,
+                    "-map",
+                    f"[{final_label}]",
+                    "-map",
+                    "0:a?",
+                ]
+            )
         elif filter_complex:
             cmd.extend(["-filter_complex", filter_complex])
 
@@ -462,6 +525,10 @@ class AiAvatarRenderService:
                 "veryfast",
                 "-crf",
                 "23",
+                "-c:a",
+                "aac",
+                "-b:a",
+                "128k",
                 "-y",
                 output_path,
             ]

apps/api/app/tasks/lipsync_tts.py

@@ -54,6 +54,160 @@ def _sign_media_url(url: str) -> str:
         return url
 
 
+def _split_script_into_sentences(script_text: str) -> list[str]:
+    """按句号/问号/感叹号/分号/换行分句（与前端 splitScriptIntoSentences 一致）."""
+    import re
+
+    text = (script_text or "").strip()
+    if not text:
+        return []
+    parts = re.split(r"[。！？!?；;\n\r]+", text)
+    return [p.strip() for p in parts if p.strip()]
+
+
+def _compute_sentence_timings(audio_data: bytes, script_text: str, total_duration: float) -> list[dict]:
+    """基于 TTS 音频的静音检测，精确计算每句文案的起止时间.
+
+    使用 ffmpeg silencedetect 检测静音段，将静音点与句子边界对齐。
+    比字数比例估算准确得多。
+
+    Args:
+        audio_data: TTS 音频二进制数据（MP3）
+        script_text: 文案全文
+        total_duration: 音频总时长（秒）
+
+    Returns:
+        list[{"index": int, "text": str, "start_time": float, "end_time": float}]
+    """
+    import re
+    import subprocess
+    import tempfile
+
+    sentences = _split_script_into_sentences(script_text)
+    if not sentences:
+        return []
+
+    # 写入临时音频文件
+    with tempfile.NamedTemporaryFile(suffix=".mp3", delete=False) as tmp:
+        tmp.write(audio_data)
+        tmp_path = tmp.name
+
+    try:
+        # 用 ffmpeg silencedetect 检测静音段
+        result = subprocess.run(
+            [
+                "ffmpeg",
+                "-i",
+                tmp_path,
+                "-af",
+                "silencedetect=noise=-25dB:d=0.3",
+                "-f",
+                "null",
+                "-",
+            ],
+            capture_output=True,
+            text=True,
+            timeout=30,
+        )
+        stderr = result.stderr or ""
+
+        # 解析静音结束时间点（silence_end: X.XXX）
+        silence_ends = []
+        for match in re.finditer(r"silence_end:\s*([\d.]+)", stderr):
+            t = float(match.group(1))
+            if 0 < t < total_duration:
+                silence_ends.append(t)
+
+        # 如果没有检测到足够的静音点，降级为字数比例估算
+        if len(silence_ends) < len(sentences) - 1:
+            logger.warning(
+                "[sentence_timings] 静音点不足(%d < %d)，降级为字数比例估算",
+                len(silence_ends),
+                len(sentences) - 1,
+            )
+            return _estimate_sentence_timings_by_chars(sentences, total_duration)
+
+        # 贪心匹配：N-1 个句子边界对应 N-1 个静音点
+        # 按时间均匀分布期望值，选择最近的静音点
+        n_boundaries = len(sentences) - 1
+        boundaries = []
+        used_indices = set()
+
+        for i in range(n_boundaries):
+            # 期望的边界位置（按句子数量均匀分布）
+            expected_pos = (i + 1) / len(sentences) * total_duration
+            # 找最近的未使用静音点
+            best_idx = None
+            best_dist = float("inf")
+            for j, t in enumerate(silence_ends):
+                if j in used_indices:
+                    continue
+                dist = abs(t - expected_pos)
+                if dist < best_dist:
+                    best_dist = dist
+                    best_idx = j
+            if best_idx is not None:
+                used_indices.add(best_idx)
+                boundaries.append(silence_ends[best_idx])
+
+        boundaries.sort()
+
+        # 构建 sentence_timings
+        timings = []
+        prev_end = 0.0
+        for i, sent in enumerate(sentences):
+            start = prev_end
+            end = boundaries[i] if i < len(boundaries) else total_duration
+            timings.append(
+                {
+                    "index": i,
+                    "text": sent,
+                    "start_time": round(start, 2),
+                    "end_time": round(end, 2),
+                }
+            )
+            prev_end = end
+
+        return timings
+
+    except Exception as exc:
+        logger.warning("[sentence_timings] 静音检测异常，降级为字数比例估算: %s", exc)
+        return _estimate_sentence_timings_by_chars(sentences, total_duration)
+    finally:
+        import os
+
+        try:
+            os.unlink(tmp_path)
+        except Exception:
+            pass
+
+
+def _estimate_sentence_timings_by_chars(sentences: list[str], total_duration: float) -> list[dict]:
+    """降级方案：按字数比例估算句子时间（与原前端逻辑一致）."""
+    if not sentences or total_duration <= 0:
+        return []
+    total_chars = sum(len(s.replace(r"\s", "")) for s in sentences)
+    if total_chars == 0:
+        return []
+
+    timings = []
+    acc = 0
+    for i, sent in enumerate(sentences):
+        chars = len(sent.replace(r"\s", ""))
+        start = (acc / total_chars) * total_duration
+        end = ((acc + chars) / total_chars) * total_duration
+        timings.append(
+            {
+                "index": i,
+                "text": sent,
+                "start_time": round(start, 2),
+                "end_time": round(end, 2),
+            }
+        )
+        acc += chars
+    return timings
+
+
 @shared_task(
     bind=True,
     name="lipsync_tts.synthesize_and_submit",
@@ -152,14 +306,14 @@ def tts_synthesize_and_submit(
             audio_data = safe_download_bytes(
                 temp_url,
                 purpose="lipsync_tts_audio",
-                allowed_mime_types=(
+                allowed_mime_types={
                     "audio/mpeg",
                     "audio/mp3",
                     "audio/wav",
                     "audio/x-wav",  # CosyVoice 部分接口返回 audio/x-wav，与 audio/wav 等价（RIFF/WAVE）
                     "audio/mp4",
                     "audio/x-m4a",
-                ),
+                },
                 timeout=60.0,
             )
             from packages.shared.storage import get_shared_storage_service
@@ -179,6 +333,64 @@ def tts_synthesize_and_submit(
 
         db.commit()
 
+        # 2.5 计算精确句子时间戳（基于 TTS 音频静音检测）
+        import os as _os
+
+        _st_tmp_path = None
+        try:
+            import subprocess as _sp
+            import tempfile as _tmpf
+
+            # 下载音频用于探测时长和静音检测
+            if isinstance(job.audio_url, str) and job.audio_url:
+                from packages.shared.url_security import safe_download_bytes as _sdl
+
+                _audio_bytes = _sdl(job.audio_url, purpose="sentence_timings", timeout=30.0)
+            else:
+                _audio_bytes = audio_data
+
+            # ffprobe 获取音频时长
+            with _tmpf.NamedTemporaryFile(suffix=".mp3", delete=False) as _atmp:
+                _atmp.write(_audio_bytes)
+                _st_tmp_path = _atmp.name
+
+            _probe_result = _sp.run(
+                [
+                    "ffprobe",
+                    "-v",
+                    "error",
+                    "-show_entries",
+                    "format=duration",
+                    "-of",
+                    "default=noprint_wrappers=1:nokey=1",
+                    _st_tmp_path,
+                ],
+                capture_output=True,
+                text=True,
+                timeout=10,
+            )
+            _audio_duration = float(_probe_result.stdout.strip()) if _probe_result.stdout.strip() else 0.0
+
+            if _audio_duration > 0:
+                _timings = _compute_sentence_timings(_audio_bytes, script_text, _audio_duration)
+                if _timings:
+                    job.sentence_timings = _timings
+                    logger.info(
+                        "[lipsync_tts] 句子时间戳已计算: job_id=%s sentences=%d duration=%.1f",
+                        job_id,
+                        len(_timings),
+                        _audio_duration,
+                    )
+            db.commit()
+        except Exception as _st_err:
+            logger.warning("[lipsync_tts] 句子时间戳计算失败（不影响主流程）: job_id=%s err=%s", job_id, _st_err)
+        finally:
+            if _st_tmp_path:
+                try:
+                    _os.unlink(_st_tmp_path)
+                except Exception:
+                    pass
+
         # 3. 签名 URL 并提交到 MediaKit（复用模块内 _sign_media_url，避免对 LipsyncService 的耦合）
         audio_url = _sign_media_url(job.audio_url)
         video_url = _sign_media_url(job.video_url)

apps/web/src/pages/ai-avatar/AiAvatarPage.tsx

@@ -241,6 +241,15 @@ const AiAvatarPage: React.FC = () => {
             if (renderTimerRef.current) clearInterval(renderTimerRef.current)
             renderTimerRef.current = null
             setRenderStatus("completed")
+            // 渲染完成后，用最终视频的封面更新前端封面配置
+            if (updated.output_cover_url) {
+              state.setCoverConfig((prev) => ({
+                ...prev,
+                mode: "auto_frame",
+                smart_cover_url: updated.output_cover_url,
+                thumbnail_url: updated.output_cover_url,
+              }))
+            }
             message.success("视频已生成并保存到成片库")
           } else if (updated.status === "failed") {
             if (renderTimerRef.current) clearInterval(renderTimerRef.current)
@@ -488,8 +497,9 @@ const AiAvatarPage: React.FC = () => {
           open={state.showBRollModal}
           onClose={() => state.setShowBRollModal(false)}
           existingSegments={state.bRollSegments}
-          scriptText={state.scriptText}
+          scriptText={state.lipsyncJob?.script_text || state.scriptText}
           outputDuration={state.lipsyncJob?.output_duration ?? 0}
+          sentenceTimings={state.lipsyncJob?.sentence_timings}
           onConfirm={state.addBRollSegment}
           onRemove={state.removeBRollSegment}
         />

apps/web/src/pages/ai-avatar/components/ModalBRollEditor.tsx

@@ -5,12 +5,12 @@
  * - 左侧：先选素材库（video 库）→ 再选该库视频素材（已被其他 segment 使用的素材
  *   标灰 + "已选择" 遮罩，pointer-events:none 防重复选择）
  * - 右侧：文案句子列表（点选对应段落，替代原数字索引框）/ 全屏 or 画中画 / 四角位置+大小
- *   （开始/结束时间已删除，按句子字数占比 × 口播总时长自动估算）
+ *   （开始/结束时间来自后端精确句子时间戳，基于 TTS 音频静音检测）
  * - 底部：已配置的画面插入列表（可删除）
  */
 import React, { useEffect, useMemo, useState } from "react"
 import { getAssets, getAssetLibraries, type AssetItem, type AssetLibraryItem } from "@/api/assets"
-import type { BRollSegment, BRollInsertMode, PipPosition } from "../types"
+import type { BRollSegment, BRollInsertMode, PipPosition, SentenceTiming } from "../types"
 import { splitScriptIntoSentences, type ScriptSentence } from "../utils/sentences"
 
 interface ModalBRollEditorProps {
@@ -18,10 +18,12 @@ interface ModalBRollEditorProps {
   onClose: () => void
   /** 当前已有的 B-roll segments（用于标灰已选素材） */
   existingSegments: BRollSegment[]
-  /** 当前文案全文（用于分句） */
+  /** 文案全文（优先使用对口型时锁定的 scriptText） */
   scriptText: string
-  /** 对口型成片总时长（秒），用于时间自动估算 */
+  /** 对口型成片总时长（秒） */
   outputDuration: number
+  /** 后端精确句子时间戳（来自 lipsyncJob.sentence_timings） */
+  sentenceTimings?: SentenceTiming[] | null
   onConfirm: (segment: BRollSegment) => void
   onRemove: (id: string) => void
 }
@@ -43,7 +45,8 @@ const ModalBRollEditor: React.FC<ModalBRollEditorProps> = ({
   onClose,
   existingSegments,
   scriptText,
-  outputDuration,
+  outputDuration: _outputDuration,
+  sentenceTimings,
   onConfirm,
   onRemove,
 }) => {
@@ -62,10 +65,10 @@ const ModalBRollEditor: React.FC<ModalBRollEditorProps> = ({
   const [pipPosition, setPipPosition] = useState<PipPosition>("top-right")
   const [pipScale, setPipScale] = useState(0.3)
 
-  /** 文案分句（⑤） */
+  /** 文案分句（使用后端精确时间戳） */
   const sentences = useMemo(
-    () => splitScriptIntoSentences(scriptText, outputDuration),
-    [scriptText, outputDuration],
+    () => splitScriptIntoSentences(scriptText, sentenceTimings),
+    [scriptText, sentenceTimings],
   )
 
   /** 已被现有 segments 占用的素材 id 集合（标灰、禁止重复选择） */
@@ -264,7 +267,7 @@ const ModalBRollEditor: React.FC<ModalBRollEditorProps> = ({
                           >
                             <span className="aa-sentence-item__idx">{sent.index + 1}</span>
                             <span className="aa-sentence-item__text">{sent.text}</span>
-                            {outputDuration > 0 && (
+                            {sent.endTime > 0 && (
                               <span className="aa-sentence-item__time">
                                 {sent.startTime.toFixed(1)}-{sent.endTime.toFixed(1)}s
                               </span>
@@ -349,7 +352,7 @@ const ModalBRollEditor: React.FC<ModalBRollEditorProps> = ({
                           selectedSentence.endTime,
                           selectedSentence.startTime + 0.5,
                         ).toFixed(1)}
-                        s （按字数自动估算）
+                        s
                       </div>
                     </>
                   ) : (

apps/web/src/pages/ai-avatar/components/PanelCoverAndGenerate.tsx

@@ -120,7 +120,7 @@ const PanelCoverAndGenerate: React.FC<PanelCoverAndGenerateProps> = ({
       wordBreak: "break-word",
       whiteSpace: "pre-wrap",
       color: titleConfig.color || "#ffffff",
-      fontSize: `${titleConfig.size}px`,
+      fontSize: `${(titleConfig.size || 28) * 0.55}px`, // 预览容器缩放，与 PanelLipsyncPreview 对齐
       fontFamily: getFontFamily(titleConfig.font),
       fontWeight: titleConfig.bold ? "bold" : "normal",
       fontStyle: titleConfig.italic ? "italic" : "normal",
@@ -153,8 +153,8 @@ const PanelCoverAndGenerate: React.FC<PanelCoverAndGenerateProps> = ({
     } else if (titleConfig.shadow) {
       style.textShadow = "0 2px 8px rgba(0,0,0,0.7), 0 0 2px rgba(0,0,0,0.5)"
     } else {
-      // 默认给轻微阴影保证白字在亮背景可读
-      style.textShadow = "0 2px 6px rgba(0,0,0,0.6)"
+      // 无描边无阴影时，不加额外效果（与后端 drawtext 对齐：无 stroke/shadow 则不加）
+      style.textShadow = "none"
     }
 
     return style

apps/web/src/pages/ai-avatar/components/PanelLipsyncPreview.tsx

@@ -14,8 +14,8 @@ interface PanelLipsyncPreviewProps {
   onRemoveBRoll: (id: string) => void
   /** 标题配置（实时叠加预览用） */
   titleConfig?: AiAvatarTitleConfig
-  /** 标题位置变更回调（拖拽结束时调用） */
-  onTitlePositionChange?: (pos: { pos_x: number; pos_y: number }) => void
+  /** 标题位置变更回调（拖拽结束时调用，发送百分比坐标 + position:"custom"） */
+  onTitlePositionChange?: (pos: { pos_x: number; pos_y: number; position: string }) => void
 }
 
 const BROLL_MODE_LABEL: Record<BRollSegment["mode"], string> = {
@@ -56,11 +56,9 @@ export function PanelLipsyncPreview({
   const titleOverlayStyle: React.CSSProperties | null = titleConfig?.title
     ? {
         position: "absolute",
-        left: "50%",
-        transform: "translateX(-50%)",
         color: titleConfig.color || "#ffffff",
         fontFamily: titleConfig.font || "思源黑体",
-        fontSize: `${(titleConfig.size || 36) * 0.55}px`, // 预览等比缩
+        fontSize: `${(titleConfig.size || 28) * 0.55}px`,
         fontWeight: titleConfig.bold ? 700 : 400,
         fontStyle: titleConfig.italic ? "italic" : "normal",
         textAlign: "center",
@@ -68,11 +66,19 @@ export function PanelLipsyncPreview({
         padding: "4px 8px",
         textShadow: titleConfig.shadow ? "0 2px 4px rgba(0,0,0,0.8)" : undefined,
         WebkitTextStroke: titleConfig.stroke ? "1.5px #000" : undefined,
-        ...(titleConfig.position === "top"
-          ? { top: 8 }
-          : titleConfig.position === "bottom"
-            ? { bottom: 8 }
-            : { top: "50%", transform: "translateX(-50%) translateY(-50%)" }),
+        ...(titleConfig.position === "custom" &&
+        titleConfig.pos_x != null &&
+        titleConfig.pos_y != null
+          ? {
+              left: `${titleConfig.pos_x}%`,
+              top: `${titleConfig.pos_y}%`,
+              transform: "translateX(-50%) translateY(-50%)",
+            }
+          : titleConfig.position === "top"
+            ? { left: "50%", top: 8, transform: "translateX(-50%)" }
+            : titleConfig.position === "bottom"
+              ? { left: "50%", bottom: 8, transform: "translateX(-50%)" }
+              : { left: "50%", top: "50%", transform: "translateX(-50%) translateY(-50%)" }),
       }
     : null
 
@@ -105,7 +111,10 @@ export function PanelLipsyncPreview({
       const rect = previewContainerRef.current.getBoundingClientRect()
       const relX = Math.max(0, Math.min(rect.width, e.clientX - rect.left))
       const relY = Math.max(0, Math.min(rect.height, e.clientY - rect.top))
-      onTitlePositionChange({ pos_x: relX, pos_y: relY })
+      // 发送百分比坐标（0-100），与后端 drawtext 百分比表达式对齐
+      const xpct = Math.round((relX / rect.width) * 1000) / 10
+      const ypct = Math.round((relY / rect.height) * 1000) / 10
+      onTitlePositionChange({ pos_x: xpct, pos_y: ypct, position: "custom" })
     }
     ;(e.currentTarget as HTMLDivElement).style.cursor = "grab"
   }

apps/web/src/pages/ai-avatar/types.ts

@@ -44,12 +44,23 @@ export interface LipsyncJob {
   status: LipsyncStatus
   progress: number
   output_video_url: string | null
-  /** 对口型成片总时长（秒），后端返回；用于 B-roll 时间自动估算（#1809 ⑥） */
+  /** 对口型成片总时长（秒），后端返回 */
+  script_text: string
   output_duration?: number
+  /** 精确句子时间戳（后端基于 TTS 音频静音检测计算） */
+  sentence_timings?: SentenceTiming[] | null
   error_message: string | null
   created_at: string
 }
 
+/* ── 句子时间戳（后端精确计算） ── */
+export interface SentenceTiming {
+  index: number
+  text: string
+  start_time: number
+  end_time: number
+}
+
 /* ── B-roll 画面插入 ── */
 export type BRollInsertMode = "fullscreen" | "pip"
 export type PipPosition = "top-left" | "top-right" | "bottom-left" | "bottom-right"
@@ -77,7 +88,7 @@ export interface AiAvatarTitleConfig {
   shadow: boolean
   color: string
   auto_subtitle: boolean
-  /** 自定义位置坐标（position=custom 时生效，像素） */
+  /** 自定义位置坐标（position=custom 时生效，百分比 0-100） */
   pos_x?: number
   pos_y?: number
 }
@@ -101,6 +112,7 @@ export interface RenderJob {
   status: RenderStatus
   progress: number
   output_video_url: string | null
+  output_cover_url: string | null
   error_message: string | null
   created_at: string
 }

apps/web/src/pages/ai-avatar/utils/contract.ts

@@ -39,7 +39,7 @@ export function buildTitleConfigPayload(cfg: AiAvatarTitleConfig): Record<string
     text,
     enabled: true,
     font: cfg.font || "思源黑体",
-    font_size: Math.round(cfg.size) || 36,
+    font_size: Math.round(cfg.size) || 28,
     font_color: cfg.color || "#ffffff",
     position,
     bold: !!cfg.bold,

apps/web/src/pages/ai-avatar/utils/sentences.ts

@@ -1,6 +1,10 @@
 /**
- * AI数字人 — 文案分句 & B-roll 时间自动估算（#1809 ⑤⑥）
+ * AI数字人 — 文案分句工具
+ *
+ * 分句规则与后端 _split_script_into_sentences 保持一致。
+ * 时间戳由后端基于 TTS 音频静音检测精确计算，前端不再做字数比例估算。
  */
+import type { SentenceTiming } from "../types"
 
 export interface ScriptSentence {
   /** 句子序号（从 0 开始，对应提交给后端的 script_segment_index） */
@@ -9,21 +13,21 @@ export interface ScriptSentence {
   text: string
   /** 句子字数（按中文/字符计，去除空白） */
   charCount: number
-  /** 累计起始字数（用于时间估算） */
+  /** 累计起始字数 */
   startChar: number
-  /** 估算的对口型视频内起始时间（秒） */
+  /** 精确起始时间（秒），来自后端 sentence_timings；无数据时为 0 */
   startTime: number
-  /** 估算的对口型视频内结束时间（秒） */
+  /** 精确结束时间（秒），来自后端 sentence_timings；无数据时为 0 */
   endTime: number
 }
 
 /**
  * 按句号/问号/感叹号/分号/换行分句（兼容中英文标点）。
- * 空文案返回空数组。时间按「该句字数 ÷ 全文总字数 × 口播总时长」线性估算。
+ * 时间戳从后端 sentence_timings 获取（精确）；若无则返回 0（由调用方降级处理）。
  */
 export function splitScriptIntoSentences(
   scriptText: string,
-  outputDuration: number,
+  sentenceTimings?: SentenceTiming[] | null,
 ): ScriptSentence[] {
   const text = (scriptText || "").trim()
   if (!text) return []
@@ -33,30 +37,25 @@ export function splitScriptIntoSentences(
     .map((part) => part.trim())
     .filter((part) => part.length > 0)
 
-  const totalChars = rawParts.reduce((sum, part) => sum + part.replace(/\s/g, "").length, 0)
-  const duration = outputDuration > 0 ? outputDuration : 0
-
   const sentences: ScriptSentence[] = []
   let accChar = 0
   rawParts.forEach((part, i) => {
     const charCount = part.replace(/\s/g, "").length
-    const startTime = duration > 0 && totalChars > 0 ? (accChar / totalChars) * duration : 0
-    const endTime =
-      duration > 0 && totalChars > 0 ? ((accChar + charCount) / totalChars) * duration : 0
+    // 从后端精确时间戳获取；无数据时返回 0
+    const timing = sentenceTimings?.[i]
+    const startTime = timing?.start_time ?? 0
+    const endTime = timing?.end_time ?? 0
+
     sentences.push({
       index: i,
       text: part,
       charCount,
       startChar: accChar,
-      startTime: round1(startTime),
-      endTime: round1(endTime),
+      startTime,
+      endTime,
     })
     accChar += charCount
   })
 
   return sentences
 }
-
-function round1(n: number): number {
-  return Math.round(n * 10) / 10
-}

packages/adapters/sqlalchemy_impl/models.py

@@ -703,6 +703,9 @@ class LipsyncJobModel(Base):
     error_message = Column(Text, nullable=False, default="")
     error_code = Column(String(100), nullable=False, default="")
 
+    # 精确句子时间戳（TTS 合成后由 silencedetect 计算，用于 B-roll 精确定位）
+    sentence_timings = Column(JSON, nullable=True)  # list[{index,text,start_time,end_time}]
+
     # 时间戳
     submitted_at = Column(DateTime, nullable=True)
     completed_at = Column(DateTime, nullable=True)

packages/domain/video_filter_builder.py

@@ -377,26 +377,30 @@ def _append_audio_concat(parts: list[str], clip_chains: list[ClipFilterChain]) -
 
 # ── 标题 drawtext 滤镜构建（#1789）─────────────────────────────────────────────
 
-# drawtext 字体搜索路径：按优先级列出常见安装位置
-# 服务器使用 Noto Sans SC（思源黑体）作为默认字体
+# drawtext 字体搜索路径：按优先级从高到低排列
+# 服务器使用 Noto Sans SC（思源黑体）作为默认字体。
+# - NotoSansSC-VF.ttf 是 worker-base.Dockerfile 中 COPY 的 VF 字体（含所有字重，无 Mono 变体），优先级最高
+# - .ttc 系列为 fonts-noto-cjk 包预装字体（Dockerfile 已删除含 Mono 变体的旧 .ttc，存在时作为 fallback）
+# - DejaVuSans 仅含拉丁字符不支持中文，已移除
 DRAWTEXT_FONT_SEARCH_PATHS: list[str] = [
+    "/usr/share/fonts/opentype/noto/NotoSansSC-VF.ttf",
     "/usr/share/fonts/opentype/noto/NotoSansCJK-Regular.ttc",
+    "/usr/share/fonts/opentype/noto/NotoSansCJK-Bold.ttc",
     "/usr/share/fonts/noto-cjk/NotoSansCJK-Regular.ttc",
     "/usr/share/fonts/google-noto-cjk/NotoSansCJK-Regular.ttc",
     "/usr/share/fonts/truetype/noto/NotoSansSC-Regular.ttf",
     "/usr/share/fonts/noto/NotoSansSC-Regular.ttf",
-    "/usr/share/fonts/truetype/dejavu/DejaVuSans.ttf",
 ]
 
-# 前端字体名 → drawtext 字体搜索关键字
+# 前端字体名 → drawtext 字体搜索关键字（匹配 DRAWTEXT_FONT_SEARCH_PATHS 中的文件名关键字）
 DRAWTEXT_FONT_MAP: dict[str, str] = {
-    "思源黑体": "NotoSansCJK",
+    "思源黑体": "NotoSansSC",
     "思源宋体": "NotoSerifCJK",
-    "苹方": "NotoSansCJK",
-    "PingFang": "NotoSansCJK",
-    "微软雅黑": "NotoSansCJK",
+    "苹方": "NotoSansSC",
+    "PingFang": "NotoSansSC",
+    "微软雅黑": "NotoSansSC",
     "楷体": "NotoSerifCJK",
-    "华康俪金黑": "NotoSansCJK",
+    "华康俪金黑": "NotoSansSC",
 }
 
 
@@ -477,13 +481,13 @@ def build_title_drawtext_filter(
 
     # ── 样式参数 ──
     font_name = title_config.get("font") or title_config.get("font_preset") or "思源黑体"
-    font_size = int(title_config.get("font_size") or title_config.get("size") or 36)
+    font_size = int(title_config.get("font_size") or title_config.get("size") or 28)
     font_color = title_config.get("font_color") or title_config.get("color") or "#ffffff"
     # 去掉 # 前缀（drawtext 用纯 hex 或颜色名）
     if font_color.startswith("#"):
         font_color = font_color[1:]
 
-    position = title_config.get("position", "top")
+    position = title_config.get("position") or "bottom"
     bold = bool(title_config.get("bold", True))
     stroke = title_config.get("stroke")
     shadow = title_config.get("shadow")
@@ -504,26 +508,30 @@ def build_title_drawtext_filter(
     params.append(f"fontsize={font_size}")
     params.append(f"fontcolor={font_color}")
 
-    # 粗体：bold 在 drawtext 中通过 font 的 Bold 变体实现
-    # 若字体有 Bold 变体可用 fontfont=bold；否则通过 borderw 模拟
-    if bold:
-        # 使用 font 参数尝试加载 Bold 变体（Noto Sans SC 有 Bold 变体文件）
-        params.append("font=bold")
+    # 粗体：drawtext 没有独立的 bold 参数，通过加大 borderw 模拟视觉粗体效果。
+    # 注意：不能使用 `font=bold`——FFmpeg drawtext 的 font 参数需要 fontconfig 能解析的
+    # 字体族名，而 "bold" 不是合法族名，会导致整个 filter_complex 解析失败（exit code 234）。
+    # 当用户未显式配置描边宽度时，bold 模式自动将 borderw 提升到 3 以模拟粗体。
 
     # 描边（borderw 需要 libfreetype 支持）
+    # 粗体无显式描边时，自动用 borderw=3 + 近色描边模拟粗体；显式 stroke 按用户配置走
+    border_width = 0
+    border_color = "000000"
     if stroke:
         if isinstance(stroke, bool):
             border_width = 2
-            border_color = "black"
+            border_color = "000000"
         elif isinstance(stroke, dict):
-            border_width = int(stroke.get("width", 2)) if stroke.get("enabled", True) else 0
-            border_color = (stroke.get("color") or "#000000").lstrip("#")
-        else:
-            border_width = 0
-            border_color = "black"
-        if border_width > 0:
-            params.append(f"borderw={border_width}")
-            params.append(f"bordercolor={border_color}")
+            if stroke.get("enabled", True):
+                border_width = int(stroke.get("width", 2))
+                border_color = (stroke.get("color") or "#000000").lstrip("#")
+    elif bold:
+        # 粗体模式且未配描边：加大描边宽度模拟粗体效果
+        border_width = 3
+        border_color = font_color  # 用字体同色描边，视觉上加粗字形而非黑边
+    if border_width > 0:
+        params.append(f"borderw={border_width}")
+        params.append(f"bordercolor={border_color}")
 
     # 阴影（shadowcolor + shadowx/y）
     if shadow:
@@ -548,8 +556,13 @@ def build_title_drawtext_filter(
         and not isinstance(pos_x, bool)
         and not isinstance(pos_y, bool)
     ):
-        params.append(f"x={int(pos_x)}")
-        params.append(f"y={int(pos_y)}")
+        # pos_x/pos_y 为百分比坐标（0-100），转换为 drawtext 表达式
+        # 例如 pos_x=50 → x=(w-text_w)*0.50（水平居中偏50%）
+        #      pos_y=30 → y=(h-text_h)*0.30
+        pct_x = max(0.0, min(100.0, float(pos_x))) / 100.0
+        pct_y = max(0.0, min(100.0, float(pos_y))) / 100.0
+        params.append(f"x=(w-text_w)*{pct_x:.4f}")
+        params.append(f"y=(h-text_h)*{pct_y:.4f}")
     else:
         # 三档预设位置：top / center / bottom
         # x 始终水平居中：(w-text_w)/2
@@ -573,7 +586,7 @@ def build_broll_overlay_filter(
     video_duration: float,
     output_width: int = DEFAULT_OUTPUT_WIDTH,
     output_height: int = DEFAULT_OUTPUT_HEIGHT,
-) -> str:
+) -> tuple[str, str | None]:
     """构建 B-roll 叠加滤镜链。
 
     支持两种模式:
@@ -581,121 +594,182 @@ def build_broll_overlay_filter(
     - pip: 在对口型视频上叠加画中画 B-roll
 
     Args:
-        b_roll_segments: B-roll 片段配置列表
+        b_roll_segments: B-roll 片段配置列表（原始顺序，决定 FFmpeg -i 输入顺序）
         video_duration: 对口型视频总时长(秒)
-        output_width: 输出宽度
-        output_height: 输出高度
+        output_width: 输出宽度（默认 1280；AI 数字人竖屏传 720）
+        output_height: 输出高度（默认 720；AI 数字人竖屏传 1280）
 
     Returns:
-        FFmpeg filter_complex 滤镜字符串片段
+        (filter_complex_str, final_label)
+        - filter_complex_str: filter_complex 片段字符串（末尾无分号）
+        - final_label: 最终输出 pad 标签名，如 "vout"；无 B-roll 时返回 None
     """
     if not b_roll_segments:
-        return ""
+        return "", None
+
+    # 建立原始列表下标 → FFmpeg 输入下标的映射：
+    # cmd 中 [0:v] 是主视频，随后按 b_roll_segments 原始顺序追加 -i，
+    # 因此第 i 个 segment 的输入是 [{i+1}:v]
+    def _input_label(seg: dict[str, Any]) -> str:
+        # seg 必须来自 b_roll_segments；通过 id() 在原列表中查找
+        for i, s in enumerate(b_roll_segments):
+            if s is seg:
+                return f"[{i + 1}:v]"
+        # fallback: 找不到时不应发生，保守返回
+        return "[1:v]"
 
     parts: list[str] = []
     sorted_segments = sorted(b_roll_segments, key=lambda s: s.get("start_time", 0))
 
-    # 按模式分组处理
+    # 按模式分组
     fullscreen_segments = [s for s in sorted_segments if s.get("mode") == "fullscreen"]
     pip_segments = [s for s in sorted_segments if s.get("mode") == "pip"]
 
+    final_label = None
+
     # ── fullscreen 模式: 切分 + concat ──
     if fullscreen_segments:
-        parts.append(_build_fullscreen_filters(fullscreen_segments, video_duration, output_width, output_height))
+        fs_filter, fs_label = _build_fullscreen_filters(
+            fullscreen_segments, b_roll_segments, video_duration, output_width, output_height, _input_label
+        )
+        parts.append(fs_filter)
+        final_label = fs_label
+    else:
+        fs_label = None
 
     # ── pip 模式: overlay 滤镜 ──
     if pip_segments:
-        for idx, seg in enumerate(pip_segments):
-            start = seg.get("start_time", 0)
-            end = seg.get("end_time", video_duration)
-            scale = seg.get("pip_scale", 0.3)
-            position = seg.get("pip_position", "bottom_right")
-
-            pip_w = int(output_width * scale)
-            pip_h = int(output_height * scale)
-
-            # 位置映射
-            pos_map = {
-                "top_left": "10:10",
-                "top_right": "W-w-10:10",
-                "bottom_left": "10:H-h-10",
-                "bottom_right": "W-w-10:H-h-10",
-                "center": "(W-w)/2:(H-h)/2",
-            }
-            pos_expr = pos_map.get(position, pos_map["bottom_right"])
-
-            broll_input_idx = len(sorted_segments)  # placeholder for input index
-            parts.append(
-                f"[{broll_input_idx + idx}:v]scale={pip_w}:{pip_h}," f"enable='between(t,{start},{end})'[pip{idx}];"
-            )
-            # overlay onto main stream
-            if idx == 0:
-                base_label = "[vout]" if fullscreen_segments else "[0:v]"
-            else:
-                base_label = f"[pip{idx - 1}]"
-            parts.append(f"{base_label}[pip{idx}]overlay={pos_expr}:enable='between(t,{start},{end})'[vout{idx}];")
+        pip_filter, pip_label = _build_pip_filters(
+            pip_segments, output_width, output_height, _input_label, base_label=fs_label
+        )
+        parts.append(pip_filter)
+        final_label = pip_label
 
     result = "".join(parts)
     # 清理末尾多余分号
     if result.endswith(";"):
         result = result[:-1]
-    return result
+    return result, final_label
 
 
 def _build_fullscreen_filters(
-    segments: list[dict[str, Any]],
+    sorted_fs_segments: list[dict[str, Any]],
+    all_segments: list[dict[str, Any]],
     video_duration: float,
     output_width: int,
     output_height: int,
-) -> str:
-    """构建 fullscreen 模式的切分 + concat 滤镜.
+    input_label_fn,
+) -> tuple[str, str]:
+    """构建 fullscreen 模式的切分 + concat 滤镜。
 
-    将对口型视频按 B-roll 时间段切分，然后用 concat 拼接 B-roll 片段。
+    将主视频按 B-roll 时间段切分，然后用 concat 拼接主视频片段和 B-roll 片段。
+
+    Returns:
+        (filter_str, final_label) 其中 final_label 是 concat 输出的 pad 标签
     """
     parts: list[str] = []
     prev_end = 0.0
 
-    for idx, seg in enumerate(segments):
+    # 注意：这里的 idx 是 sorted_fs_segments 中的下标；
+    # 实际 FFmpeg 输入下标必须通过 input_label_fn 查询
+    for idx, seg in enumerate(sorted_fs_segments):
         start = seg.get("start_time", 0)
         end = seg.get("end_time", video_duration)
 
-        # 保持原视频片段（B-roll 之前的部分）
+        # 主视频片段（B-roll 之前）
         if prev_end < start:
             parts.append(f"[0:v]trim=start={prev_end}:end={start},setpts=PTS-STARTPTS[main{idx}];")
 
-        # B-roll 片段：缩放至目标分辨率
+        # B-roll 片段：缩放到输出分辨率并裁到对应时长
+        in_lbl = input_label_fn(seg)
         parts.append(
-            f"[{idx + 1}:v]scale={output_width}:{output_height}"
+            f"{in_lbl}scale={output_width}:{output_height}"
             f":force_original_aspect_ratio=decrease,"
             f"pad={output_width}:{output_height}:(ow-iw)/2:(oh-ih)/2,"
             f"trim=start=0:end={end - start},setpts=PTS-STARTPTS[br{idx}];"
         )
         prev_end = end
 
-    # 尾部片段
+    # 尾部主视频片段
     if prev_end < video_duration:
-        last_idx = len(segments)
+        last_idx = len(sorted_fs_segments)
         parts.append(f"[0:v]trim=start={prev_end}:end={video_duration},setpts=PTS-STARTPTS[main{last_idx}];")
 
     # concat 所有片段
-    segment_labels = []
-    for idx in range(len(segments)):
-        start = segments[idx].get("start_time", 0)
-        if (idx == 0 and segments[0].get("start_time", 0) > 0) or idx > 0:
-            prev_end_prev = segments[idx - 1].get("end_time", 0) if idx > 0 else 0
-            if prev_end_prev < start:
-                segment_labels.append(f"[main{idx}]")
+    segment_labels: list[str] = []
+    for idx, seg in enumerate(sorted_fs_segments):
+        start = seg.get("start_time", 0)
+        # 每段 B-roll 之前是否有主视频片段？
+        has_main_before = (idx == 0 and start > 0) or (
+            idx > 0 and sorted_fs_segments[idx - 1].get("end_time", 0) < start
+        )
+        if has_main_before:
+            segment_labels.append(f"[main{idx}]")
         segment_labels.append(f"[br{idx}]")
-
     if prev_end < video_duration:
-        segment_labels.append(f"[main{len(segments)}]")
+        segment_labels.append(f"[main{len(sorted_fs_segments)}]")
 
+    final_lbl = "vout_fs"
     n = len(segment_labels)
     if n > 0:
         concat_inputs = "".join(segment_labels)
-        parts.append(f"{concat_inputs}concat=n={n}:v=1:a=0[vout];")
+        parts.append(f"{concat_inputs}concat=n={n}:v=1:a=0[{final_lbl}];")
 
-    return "".join(parts)
+    return "".join(parts), final_lbl
+
+
+def _build_pip_filters(
+    pip_segments: list[dict[str, Any]],
+    output_width: int,
+    output_height: int,
+    input_label_fn,
+    base_label: str | None,
+) -> tuple[str, str]:
+    """构建 PIP（画中画）overlay 滤镜链。
+
+    Args:
+        pip_segments: 按时间排序的 pip 片段
+        output_width: 输出宽度
+        output_height: 输出高度
+        input_label_fn: 片段 → 输入标签的映射函数
+        base_label: 前序滤镜链输出的标签（如 fullscreen 的 vout_fs），为 None 则基于 [0:v]
+
+    Returns:
+        (filter_str, final_label)
+    """
+    parts: list[str] = []
+    cur_label = base_label  # 当前叠加到的标签
+
+    pos_map = {
+        "top_left": "10:10",
+        "top_right": "W-w-10:10",
+        "bottom_left": "10:H-h-10",
+        "bottom_right": "W-w-10:H-h-10",
+        "center": "(W-w)/2:(H-h)/2",
+    }
+
+    for idx, seg in enumerate(pip_segments):
+        start = seg.get("start_time", 0)
+        end = seg.get("end_time", 0)
+        scale = seg.get("pip_scale", 0.3)
+        position = seg.get("pip_position", "bottom_right")
+        pos_expr = pos_map.get(position, pos_map["bottom_right"])
+
+        pip_w = max(1, int(output_width * scale))
+        pip_h = max(1, int(output_height * scale))
+        enable_expr = f"enable='between(t,{start},{end})'"
+
+        in_lbl = input_label_fn(seg)
+        pip_scaled = f"pip{idx}"
+        parts.append(f"{in_lbl}scale={pip_w}:{pip_h},{enable_expr}[{pip_scaled}];")
+
+        # overlay onto the current base
+        base = f"[{cur_label}]" if cur_label else "[0:v]"
+        out_lbl = f"vout_pip{idx}" if idx < len(pip_segments) - 1 else "vout"
+        parts.append(f"{base}[{pip_scaled}]overlay={pos_expr}:{enable_expr}[{out_lbl}];")
+        cur_label = out_lbl
+
+    return "".join(parts), cur_label or "vout"
 
 
 def build_cover_extract_command(

tests/unit/test_ai_avatar_render_routes.py

@@ -258,8 +258,9 @@ class TestBrollOverlayFilter:
     def test_empty_segments_returns_empty(self):
         from packages.domain.video_filter_builder import build_broll_overlay_filter
 
-        result = build_broll_overlay_filter([], 30.0)
+        result, label = build_broll_overlay_filter([], 30.0)
         assert result == ""
+        assert label is None
 
     def test_pip_mode_generates_overlay(self):
         from packages.domain.video_filter_builder import build_broll_overlay_filter
@@ -275,8 +276,9 @@ class TestBrollOverlayFilter:
                 "pip_scale": 0.3,
             }
         ]
-        result = build_broll_overlay_filter(segments, 30.0)
+        result, label = build_broll_overlay_filter(segments, 30.0)
         assert "overlay" in result or "scale=" in result
+        assert label == "vout"
 
     def test_fullscreen_mode_generates_concat(self):
         from packages.domain.video_filter_builder import build_broll_overlay_filter
@@ -290,8 +292,9 @@ class TestBrollOverlayFilter:
                 "end_time": 10.0,
             }
         ]
-        result = build_broll_overlay_filter(segments, 30.0)
+        result, label = build_broll_overlay_filter(segments, 30.0)
         assert "trim" in result or "concat" in result
+        assert label == "vout_fs"
 
     def test_cover_extract_command(self):
         from packages.domain.video_filter_builder import build_cover_extract_command

tests/unit/test_sentence_timings.py

@@ -0,0 +1,171 @@
+"""Tests for sentence timing functions in lipsync_tts."""
+
+import os
+import subprocess
+import tempfile
+import unittest
+from unittest.mock import MagicMock, patch
+
+from apps.api.app.tasks.lipsync_tts import (
+    _compute_sentence_timings,
+    _estimate_sentence_timings_by_chars,
+    _split_script_into_sentences,
+)
+
+
+class TestSplitScriptIntoSentences(unittest.TestCase):
+    """Tests for _split_script_into_sentences."""
+
+    def test_empty_string(self):
+        self.assertEqual(_split_script_into_sentences(""), [])
+
+    def test_none(self):
+        self.assertEqual(_split_script_into_sentences(None), [])
+
+    def test_whitespace_only(self):
+        self.assertEqual(_split_script_into_sentences("   \n  "), [])
+
+    def test_single_sentence(self):
+        self.assertEqual(_split_script_into_sentences("你好世界。"), ["你好世界"])
+
+    def test_multiple_sentences_chinese(self):
+        result = _split_script_into_sentences("第一句。第二句！第三句？")
+        self.assertEqual(result, ["第一句", "第二句", "第三句"])
+
+    def test_english_punctuation(self):
+        result = _split_script_into_sentences("Hello World! How are you?")
+        self.assertEqual(result, ["Hello World", "How are you"])
+
+    def test_semicolons(self):
+        result = _split_script_into_sentences("第一部分；第二部分;第三部分")
+        self.assertEqual(result, ["第一部分", "第二部分", "第三部分"])
+
+    def test_newlines(self):
+        result = _split_script_into_sentences("第一行\n第二行\n第三行")
+        self.assertEqual(result, ["第一行", "第二行", "第三行"])
+
+    def test_no_trailing_punctuation(self):
+        result = _split_script_into_sentences("没有标点的句子")
+        self.assertEqual(result, ["没有标点的句子"])
+
+
+class TestEstimateSentenceTimingsByChars(unittest.TestCase):
+    """Tests for _estimate_sentence_timings_by_chars."""
+
+    def test_empty_sentences(self):
+        self.assertEqual(_estimate_sentence_timings_by_chars([], 10.0), [])
+
+    def test_zero_duration(self):
+        self.assertEqual(_estimate_sentence_timings_by_chars(["hello"], 0), [])
+
+    def test_negative_duration(self):
+        self.assertEqual(_estimate_sentence_timings_by_chars(["hello"], -5.0), [])
+
+    def test_single_sentence(self):
+        result = _estimate_sentence_timings_by_chars(["hello"], 10.0)
+        self.assertEqual(len(result), 1)
+        self.assertAlmostEqual(result[0]["start_time"], 0.0)
+        self.assertAlmostEqual(result[0]["end_time"], 10.0)
+
+    def test_two_equal_sentences(self):
+        result = _estimate_sentence_timings_by_chars(["你好", "世界"], 10.0)
+        self.assertEqual(len(result), 2)
+        self.assertAlmostEqual(result[0]["start_time"], 0.0)
+        self.assertAlmostEqual(result[0]["end_time"], 5.0)
+        self.assertAlmostEqual(result[1]["start_time"], 5.0)
+        self.assertAlmostEqual(result[1]["end_time"], 10.0)
+
+    def test_unequal_char_distribution(self):
+        result = _estimate_sentence_timings_by_chars(["ABCD", "EF"], 9.0)
+        self.assertEqual(len(result), 2)
+        self.assertAlmostEqual(result[0]["start_time"], 0.0)
+        self.assertAlmostEqual(result[0]["end_time"], 6.0)  # 4/6 * 9 = 6
+        self.assertAlmostEqual(result[1]["start_time"], 6.0)
+        self.assertAlmostEqual(result[1]["end_time"], 9.0)
+
+    def test_timing_structure(self):
+        result = _estimate_sentence_timings_by_chars(["句子一", "句子二"], 6.0)
+        for item in result:
+            self.assertIn("index", item)
+            self.assertIn("text", item)
+            self.assertIn("start_time", item)
+            self.assertIn("end_time", item)
+
+
+class TestComputeSentenceTimings(unittest.TestCase):
+    """Tests for _compute_sentence_timings."""
+
+    def test_empty_script_returns_empty(self):
+        self.assertEqual(_compute_sentence_timings(b"fake_audio", "", 10.0), [])
+
+    def test_none_script_returns_empty(self):
+        self.assertEqual(_compute_sentence_timings(b"fake_audio", None, 10.0), [])
+
+    @patch("os.unlink")
+    @patch.object(tempfile, "NamedTemporaryFile")
+    @patch.object(subprocess, "run")
+    def test_silence_detection_insufficient_fallback(self, mock_run, mock_tmpfile, mock_unlink):
+        """When silence detection finds too few points, fallback to char estimation."""
+        mock_run.return_value = MagicMock(stderr="", returncode=0)
+        mock_tmp = MagicMock()
+        mock_tmp.name = "/tmp/fake.mp3"
+        mock_tmp.__enter__ = MagicMock(return_value=mock_tmp)
+        mock_tmp.__exit__ = MagicMock(return_value=False)
+        mock_tmpfile.return_value = mock_tmp
+
+        result = _compute_sentence_timings(b"fake_audio", "第一句。第二句。第三句。", 10.0)
+
+        # Should fallback to char estimation with 3 sentences
+        self.assertEqual(len(result), 3)
+        self.assertAlmostEqual(result[0]["start_time"], 0.0)
+
+    @patch("os.unlink")
+    @patch.object(tempfile, "NamedTemporaryFile")
+    @patch.object(subprocess, "run")
+    def test_silence_detection_with_enough_points(self, mock_run, mock_tmpfile, mock_unlink):
+        """When silence detection finds enough points, use them for boundaries."""
+        mock_run.return_value = MagicMock(
+            stderr="[silencedetect] silence_end: 3.5 | silence_duration: 0.4\n"
+            "[silencedetect] silence_end: 7.0 | silence_duration: 0.3\n",
+            returncode=0,
+        )
+        mock_tmp = MagicMock()
+        mock_tmp.name = "/tmp/fake.mp3"
+        mock_tmp.__enter__ = MagicMock(return_value=mock_tmp)
+        mock_tmp.__exit__ = MagicMock(return_value=False)
+        mock_tmpfile.return_value = mock_tmp
+
+        result = _compute_sentence_timings(b"fake_audio", "第一句。第二句。第三句。", 10.0)
+
+        self.assertEqual(len(result), 3)
+        self.assertAlmostEqual(result[0]["start_time"], 0.0)
+        self.assertAlmostEqual(result[0]["end_time"], 3.5)
+        self.assertAlmostEqual(result[1]["start_time"], 3.5)
+        self.assertAlmostEqual(result[1]["end_time"], 7.0)
+        self.assertAlmostEqual(result[2]["start_time"], 7.0)
+        self.assertAlmostEqual(result[2]["end_time"], 10.0)
+
+    @patch("os.unlink")
+    @patch.object(tempfile, "NamedTemporaryFile")
+    @patch.object(subprocess, "run")
+    def test_ffmpeg_exception_fallback(self, mock_run, mock_tmpfile, mock_unlink):
+        """When ffmpeg raises an exception, fallback to char estimation."""
+        mock_run.side_effect = Exception("ffmpeg not found")
+        mock_tmp = MagicMock()
+        mock_tmp.name = "/tmp/fake.mp3"
+        mock_tmp.__enter__ = MagicMock(return_value=mock_tmp)
+        mock_tmp.__exit__ = MagicMock(return_value=False)
+        mock_tmpfile.return_value = mock_tmp
+
+        result = _compute_sentence_timings(b"fake_audio", "句子一。句子二。", 6.0)
+
+        # Should fallback to char estimation
+        self.assertEqual(len(result), 2)
+        self.assertAlmostEqual(result[0]["start_time"], 0.0)
+        self.assertAlmostEqual(result[0]["end_time"], 3.0)
+        self.assertAlmostEqual(result[1]["start_time"], 3.0)
+        self.assertAlmostEqual(result[1]["end_time"], 6.0)
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/unit/test_video_filter_builder.py

@@ -902,9 +902,10 @@ class TestResolveFontPath(unittest.TestCase):
 
     @patch("os.path.isfile")
     def test_unknown_font_fallback(self, mock_isfile):
-        mock_isfile.side_effect = lambda p: "DejaVu" in p
+        # DejaVuSans 已从 fallback 列表移除（不支持 CJK），用 VF 路径模拟
+        mock_isfile.side_effect = lambda p: "NotoSansSC-VF" in p
         result = _resolve_font_path("UnknownFont")
-        self.assertIn("DejaVu", result)
+        self.assertIn("NotoSansSC-VF", result)
 
     @patch("os.path.isfile")
     def test_no_fonts_available(self, mock_isfile):
@@ -927,9 +928,11 @@ class TestResolveFontPath(unittest.TestCase):
 
     @patch("os.path.isfile")
     def test_font_fallback_skips_nonexistent(self, mock_isfile):
-        mock_isfile.side_effect = lambda p: "DejaVu" in p
+        # 所有中文字体路径都不存在时，fallback 返回第一个存在的文件；
+        # DejaVuSans 已从列表移除（不支持 CJK），使用 VF 字体路径模拟存在文件
+        mock_isfile.side_effect = lambda p: "NotoSansSC-VF" in p
         result = _resolve_font_path("不存在字体")
-        self.assertIn("DejaVu", result)
+        self.assertIn("NotoSansSC-VF", result)
 
 
 class TestDrawtextFontFileIncluded(unittest.TestCase):
@@ -1028,6 +1031,14 @@ class TestDrawtextBoldFalse(unittest.TestCase):
         self.assertIsNotNone(result)
         self.assertNotIn("font=bold", result)
 
+    def test_bold_true_does_not_use_font_bold_param(self):
+        """粗体模式不得使用 `font=bold`——该参数无效，会导致 filter_complex 解析失败（exit 234）。"""
+        result = build_title_drawtext_filter({"text": "标题", "bold": True})
+        self.assertIsNotNone(result)
+        self.assertNotIn("font=bold", result)
+        # 粗体应通过 borderw 实现
+        self.assertIn("borderw=", result)
+
 
 class TestDrawtextPositionBranches(unittest.TestCase):
     """位置相关分支覆盖。"""
@@ -1054,12 +1065,23 @@ class TestDrawtextPositionBranches(unittest.TestCase):
         self.assertIn("y=h-text_h-50", result)
 
     @patch("packages.domain.video_filter_builder._resolve_font_path")
-    def test_position_custom_with_float_coords(self, mock_font):
+    def test_position_custom_with_percentage_coords(self, mock_font):
+        """自定义位置：百分比坐标转换为 drawtext 表达式."""
+        mock_font.return_value = ""
+        # pos_x=50, pos_y=30 → x=(w-text_w)*0.5000, y=(h-text_h)*0.3000
+        result = build_title_drawtext_filter({"text": "标题", "position": "custom", "pos_x": 50, "pos_y": 30})
+        self.assertIsNotNone(result)
+        self.assertIn("x=(w-text_w)*0.5000", result)
+        self.assertIn("y=(h-text_h)*0.3000", result)
+
+    @patch("packages.domain.video_filter_builder._resolve_font_path")
+    def test_position_custom_clamped_to_100(self, mock_font):
+        """自定义位置：超过100的坐标被截断到100%."""
         mock_font.return_value = ""
         result = build_title_drawtext_filter({"text": "标题", "position": "custom", "pos_x": 100.7, "pos_y": 200.3})
         self.assertIsNotNone(result)
-        self.assertIn("x=100", result)
-        self.assertIn("y=200", result)
+        self.assertIn("x=(w-text_w)*1.0000", result)
+        self.assertIn("y=(h-text_h)*1.0000", result)
 
     @patch("packages.domain.video_filter_builder._resolve_font_path")
     def test_position_custom_bool_coords_fallback(self, mock_font):