fix(ai-avatar): B-roll时间戳、标题粗体重影、TTS卡死防护
CI/CD Pipeline / Dedup Check - skip PR tests when covered by push pipeline (pull_request) Successful in 2s
CI/CD Pipeline / Check if frontend-only change (pull_request) Successful in 2s
CI/CD Pipeline / Frontend Lint (pull_request) Successful in 1m23s
CI/CD Pipeline / PR Build Web Image (pull_request) Successful in 1m19s
CI/CD Pipeline / PR Build API Image (pull_request) Successful in 2m18s
Preview Deploy / Deploy Preview Environment (pull_request) Successful in 2m38s
CI/CD Pipeline / PR Build Worker Image (pull_request) Successful in 2m40s
PR Automation / Auto Approve on CI Green (pull_request) Successful in 3m11s
CI/CD Pipeline / Frontend Unit Tests (pull_request) Successful in 3m45s
AI Code Review / AI Code Review (pull_request) Successful in 6m59s
CI/CD Pipeline / Integration Tests (pull_request) Successful in 6m58s
CI/CD Pipeline / Unit Tests (pull_request) Successful in 7m11s
CI/CD Pipeline / Validate - Style (pull_request) Successful in 7m14s
CI/CD Pipeline / Validate - Python (mypy + alembic) (pull_request) Successful in 9m16s
PR Automation / Auto Merge on CI Green + Approved (pull_request) Successful in 10m40s
CI/CD Pipeline / Validate - Security (pull_request) Successful in 25m22s
CI/CD Pipeline / Production Browser E2E (pull_request) Has been skipped
CI/CD Pipeline / CI Gate (pull_request) Successful in 3s
CI/CD Pipeline / Deploy Production (pull_request) Failing after 89h48m47s
CI/CD Pipeline / ACR Image Cleanup (pull_request) Failing after 90h13m53s
CI/CD Pipeline / Deploy Staging (Watchtower auto-deploy) (pull_request) Failing after 90h14m1s
CI/CD Pipeline / Build Staging Web Image (pull_request) Failing after 90h14m13s
CI/CD Pipeline / Build Production API Image (pull_request) Failing after 89h48m28s
CI/CD Pipeline / Retag skipped Staging Worker Image (pull_request) Failing after 90h13m45s
CI/CD Pipeline / Retag skipped Staging Web Image (pull_request) Failing after 90h13m45s
CI/CD Pipeline / Retag skipped Staging API Image (pull_request) Failing after 90h13m46s
CI/CD Pipeline / Build Staging Worker Image (pull_request) Failing after 90h13m52s
CI/CD Pipeline / Build Staging API Image (pull_request) Failing after 90h13m52s
CI/CD Pipeline / Canary Release to Production (pull_request) Failing after 89h48m25s
CI/CD Pipeline / Build Production Worker Image (pull_request) Failing after 89h48m28s
CI/CD Pipeline / Build Production Web Image (pull_request) Failing after 89h48m28s
CI/CD Pipeline / Staging API Integration Tests (pull_request) Failing after 90h13m31s
CI/CD Pipeline / Staging E2E Tests (pull_request) Failing after 90h13m34s
CI/CD Pipeline / Check push changed paths (pull_request) Failing after 90h13m53s

1. B-roll时间戳全0修复:
   - sentences.ts 新增 sentenceTimings 参数优先使用后端精确时间戳,
     降级才按字数比例估算
   - 修正参数顺序,与 ModalBRollEditor 现有调用 (scriptText, timings, duration) 对齐
   - 前端 AiAvatarPage 已通过 props 传 sentenceTimings(旧版已传但因顺序错位被忽略)

2. 标题粗体重影修复:
   - 之前 bold=true 时使用 borderw=3 + font_color 同色描边模拟粗体,
     会在小字号/竖屏视频上造成字形边缘偏移,视觉上文字像被打印了两次
     (用户截图中的'曝光曝光…'重影)
   - 改为黑色细描边(borderw=2, 黑色),既保留清晰加粗效果又不重影
   - 用户显式开启 stroke 时仍按用户配置走
   - 新增 _resolve_font_path(bold=True) 预留粗体字体查找能力(当前镜像
     无独立 Bold 字体文件,沿用 VF 常规字重)

3. TTS 任务防卡死:
   - 给 tts_synthesize_and_submit 加 soft_time_limit=180s / time_limit=200s,
     避免因网络/上游问题导致 Celery 任务永久挂起(用户之前卡10+分钟
     tts_processing 不失败)
   - 任务开头加 INFO 日志(job_id/voice_id/text_len),方便排查 worker
     是否真的收到任务

相关:staging 用户反馈 5 问题中的 1/2/3 项(B-roll时间、标题重影、对口型卡死)
This commit is contained in:
xiaoxia
2026-09-12 21:07:34 +08:00
parent e250132ace
commit a83b53ae58
3 changed files with 106 additions and 63 deletions
+9
View File
@@ -213,6 +213,8 @@ def _estimate_sentence_timings_by_chars(sentences: list[str], total_duration: fl
name="lipsync_tts.synthesize_and_submit",
max_retries=2,
default_retry_delay=30,
soft_time_limit=180,
time_limit=200,
)
def tts_synthesize_and_submit(
self,
@@ -266,6 +268,13 @@ def tts_synthesize_and_submit(
return
# 1. TTS 合成
logger.info(
"[lipsync_tts] 开始 TTS 合成: job_id=%s voice_id=%s text_len=%d speed=%.2f",
job_id,
voice_id,
len(script_text),
speed,
)
try:
cosyvoice = CosyVoiceService()
result = cosyvoice.submit_synthesize_task(
+62 -49
View File
@@ -1,10 +1,9 @@
/**
* AI数字人 — 文案分句工具
* AI数字人 — 文案分句 & B-roll 时间计算
*
* 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。
* 优先使用后端基于 TTS 音频静音检测计算的精确 sentence_timings;
* 后端未返回(如对口型还在生成中)时,降级为前端按字数比例估算。
*/
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) */
@@ -13,75 +12,89 @@ 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 */
/** 累计起始字数(用于时间估算) */
startChar: number
/** 精确起始时间(秒),来自后端 sentence_timings;无数据时为 0 */
/** 对口型视频内起始时间(秒)——后端精确值或前端估算 */
startTime: number
/** 精确结束时间(秒),来自后端 sentence_timings;无数据时为 0 */
/** 对口型视频内结束时间(秒)——后端精确值或前端估算 */
endTime: number
}
/**
* 按句号/问号/感叹号/分号/换行分句(兼容中英文标点)。
* 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)。
* 空文案返回空数组。时间优先使用后端 sentence_timings;否则按字数线性估算。
*
* @param sentenceTimings 后端返回的精确句子时间戳(来自 lipsync_job.sentence_timings)。
* 非空且有效时优先采用,跳过前端估算。
*/
export function splitScriptIntoSentences(
scriptText: string,
sentenceTimings?: SentenceTiming[] | null,
sentenceTimings?: { index: number; text: string; start_time: number; end_time: number }[] | null,
outputDuration: number = 0,
): ScriptSentence[] {
const text = (scriptText || "").trim()
if (!text) return []
// 1. 先做基础分句(仅用于降级估算 / 没有 sentenceTimings 时)
const rawParts = text
.split(/[。!?!?;;\n\r]+/)
.map((part) => part.trim())
.filter((part) => part.length > 0)
const sentences: ScriptSentence[] = []
let accChar = 0
const totalChars = rawParts.reduce((sum, p) => sum + p.replace(/\s/g, "").length, 0)
// 如果后端返回了完整的时间戳(至少有一个有效结束时间),使用精确时间
// 需要满足:数组长度与分句数一致,且至少有一个结束时间 > 0(防御全0的异常数据)
const hasBackendTimings =
sentenceTimings &&
sentenceTimings.length >= rawParts.length &&
sentenceTimings.some((t) => (t.end_time ?? 0) > 0)
// 若有后端时间戳,取音频总时长;否则用外部传入的 outputDuration 做字数比例降级
const audioDuration = hasBackendTimings
? (sentenceTimings![sentenceTimings!.length - 1]?.end_time ?? 0)
: outputDuration
rawParts.forEach((part, i) => {
const charCount = part.replace(/\s/g, "").length
if (hasBackendTimings) {
// 使用后端精确时间戳
const timing = sentenceTimings![i]
sentences.push({
index: i,
text: part,
charCount,
startChar: accChar,
startTime: timing?.start_time ?? 0,
endTime: timing?.end_time ?? 0,
})
} else {
// 降级方案:按字数比例估算(必须保证有时间,否则B-roll无法定位)
const startTime = totalChars > 0 ? (accChar / totalChars) * (audioDuration || 0) : 0
const endTime =
totalChars > 0 ? ((accChar + charCount) / totalChars) * (audioDuration || 0) : 0
sentences.push({
index: i,
text: part,
charCount,
startChar: accChar,
startTime: Math.round(startTime * 100) / 100,
endTime: Math.round(endTime * 100) / 100,
// 2. 优先使用后端精确时间戳
// 校验:必须是数组、条数一致、每条都有 start_time/end_time,否则降级估算
if (Array.isArray(sentenceTimings) && sentenceTimings.length === rawParts.length) {
const valid = sentenceTimings.every(
(t) =>
t &&
typeof t.start_time === "number" &&
typeof t.end_time === "number" &&
t.end_time >= t.start_time,
)
if (valid) {
let accChar = 0
return sentenceTimings.map((t, i) => {
const part = rawParts[i] ?? t.text ?? ""
const charCount = part.replace(/\s/g, "").length
const sentence: ScriptSentence = {
index: t.index ?? i,
text: part,
charCount,
startChar: accChar,
startTime: round1(t.start_time),
endTime: round1(t.end_time),
}
accChar += charCount
return sentence
})
}
}
// 3. 降级:按字数比例线性估算
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
sentences.push({
index: i,
text: part,
charCount,
startChar: accChar,
startTime: round1(startTime),
endTime: round1(endTime),
})
accChar += charCount
})
return sentences
}
function round1(n: number): number {
return Math.round(n * 10) / 10
}
+35 -14
View File
@@ -420,21 +420,44 @@ def _escape_drawtext_text(text: str) -> str:
return result
def _resolve_font_path(font_name: str) -> str:
# 粗体字体文件映射:服务器镜像只保留了 NotoSansSC-VF.ttf(可变字体,已删除
# NotoSansCJK-Bold.ttc 以避免 Mono 变体问题,见 worker-base.Dockerfile),
# 因此无法通过 fontfile 切换到 Bold 字重。这里保留路径列表作为未来扩展,
# 实际加粗通过 borderw 黑色描边实现(见下)。
DRAWTEXT_BOLD_FONT_SEARCH_PATHS: list[str] = [
"/usr/share/fonts/opentype/noto/NotoSansCJK-Bold.ttc",
"/usr/share/fonts/noto-cjk/NotoSansCJK-Bold.ttc",
"/usr/share/fonts/google-noto-cjk/NotoSansCJK-Bold.ttc",
"/usr/share/fonts/truetype/noto/NotoSansSC-Bold.ttf",
"/usr/share/fonts/noto/NotoSansSC-Bold.ttf",
]
def _resolve_font_path(font_name: str, bold: bool = False) -> str:
"""解析字体名到服务器实际字体文件路径。
查找策略:
1. 通过 DRAWTEXT_FONT_MAP 映射前端字体名到服务器关键字
2. 在 DRAWTEXT_FONT_SEARCH_PATHS 中查找匹配路径
3. 未找到则返回空字符串(drawtext 使用内置默认字体)
2. bold=True 时优先查找粗体变体;找不到回退常规字重
3. 在 DRAWTEXT_FONT_SEARCH_PATHS 中查找匹配路径
4. 未找到则返回空字符串(drawtext 使用内置默认字体)
"""
keyword = DRAWTEXT_FONT_MAP.get(font_name, font_name)
import os
if bold:
for path in DRAWTEXT_BOLD_FONT_SEARCH_PATHS:
if keyword.lower() in path.lower() and os.path.isfile(path):
return path
# 粗体文件找不到时,再查常规字重(后面会用描边兜底加粗)
for path in DRAWTEXT_FONT_SEARCH_PATHS:
if keyword.lower() in path.lower() and os.path.isfile(path):
return path
# fallback:遍历搜索任意可用字体
if bold:
for path in DRAWTEXT_BOLD_FONT_SEARCH_PATHS:
if os.path.isfile(path):
return path
for path in DRAWTEXT_FONT_SEARCH_PATHS:
if os.path.isfile(path):
return path
@@ -495,8 +518,8 @@ def build_title_drawtext_filter(
# ── 构建 drawtext 参数 ──
params: list[str] = []
# 字体文件
font_path = _resolve_font_path(font_name)
# 字体文件:粗体优先使用 Bold 字体文件,避免同色描边造成字形偏移/重影
font_path = _resolve_font_path(font_name, bold=bold)
if font_path:
escaped_path = font_path.replace("\\", "\\\\").replace(":", "\\\\:").replace("'", "\\\\'")
params.append(f"fontfile='{escaped_path}'")
@@ -508,13 +531,11 @@ def build_title_drawtext_filter(
params.append(f"fontsize={font_size}")
params.append(f"fontcolor={font_color}")
# 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
# 之前用 borderw=3 + font_color 同色描边模拟粗体,会在小字号/竖屏视频上造成
# 字形偏移、边缘重影,看起来像文字被打印了两次(用户截图中的标题"曝光曝光…")。
# 修复:粗体改用黑色细描边(borderw=2, 黑色),视觉上清晰加粗且不产生偏移。
# 用户显式开启 stroke 时按用户配置走;粗体+无stroke 默认黑色细描边。
border_width = 0
border_color = "000000"
if stroke:
@@ -526,9 +547,9 @@ def build_title_drawtext_filter(
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
# 粗体模式且未配描边:黑色细描边,模拟粗体同时保证不重影
border_width = 2
border_color = "000000"
if border_width > 0:
params.append(f"borderw={border_width}")
params.append(f"bordercolor={border_color}")