Files
xiaoxia-saas/apps/api/app/tasks/lipsync_tts.py
T
xiaoxia-agent 38e4dfd628
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 / PR Build API Image (pull_request) Successful in 45s
Preview Deploy / Deploy Preview Environment (pull_request) Successful in 1m26s
CI/CD Pipeline / PR Build Worker Image (pull_request) Successful in 1m52s
PR Automation / Auto Approve on CI Green (pull_request) Successful in 3m0s
CI/CD Pipeline / Validate - Style (pull_request) Successful in 3m58s
CI/CD Pipeline / Validate - Python (mypy + alembic) (pull_request) Successful in 4m9s
CI/CD Pipeline / Integration Tests (pull_request) Successful in 4m39s
AI Code Review / AI Code Review (pull_request) Successful in 6m46s
CI/CD Pipeline / Validate - Security (pull_request) Successful in 8m17s
CI/CD Pipeline / Unit Tests (pull_request) Successful in 8m46s
CI/CD Pipeline / Production Browser E2E (pull_request) Has been skipped
CI/CD Pipeline / CI Gate (pull_request) Successful in 1s
PR Automation / Auto Merge on CI Green + Approved (pull_request) Successful in 6m25s
Preview Cleanup / Cleanup Preview Environment (pull_request) Successful in 3m10s
ACR Cleanup / ACR Image Cleanup (pull_request_target) Successful in 3m20s
CI/CD Pipeline / Deploy Production (pull_request) Failing after 35h4m20s
CI/CD Pipeline / Build Production API Image (pull_request) Failing after 35h4m24s
CI/CD Pipeline / Retag skipped Staging Web Image (pull_request) Failing after 35h13m8s
CI/CD Pipeline / Frontend Unit Tests (pull_request) Failing after 35h13m9s
CI/CD Pipeline / Build Staging API Image (pull_request) Failing after 35h13m13s
CI/CD Pipeline / PR Build Web Image (pull_request) Failing after 35h12m36s
CI/CD Pipeline / Build Staging Worker Image (pull_request) Failing after 35h12m40s
CI/CD Pipeline / Build Staging Web Image (pull_request) Failing after 35h12m40s
CI/CD Pipeline / Build Production Worker Image (pull_request) Failing after 35h3m51s
CI/CD Pipeline / Build Production Web Image (pull_request) Failing after 35h3m51s
CI/CD Pipeline / ACR Image Cleanup (pull_request) Failing after 35h12m32s
CI/CD Pipeline / Deploy Staging (Watchtower auto-deploy) (pull_request) Failing after 35h12m34s
CI/CD Pipeline / Retag skipped Staging Worker Image (pull_request) Failing after 35h12m34s
CI/CD Pipeline / Retag skipped Staging API Image (pull_request) Failing after 35h12m35s
CI/CD Pipeline / Canary Release to Production (pull_request) Failing after 35h3m47s
CI/CD Pipeline / Staging API Integration Tests (pull_request) Failing after 35h12m32s
CI/CD Pipeline / Frontend Lint (pull_request) Failing after 35h12m39s
CI/CD Pipeline / Check push changed paths (pull_request) Failing after 35h12m41s
CI/CD Pipeline / Staging E2E Tests (pull_request) Failing after 35h48m4s
feat(#1898): TTS 参数适配 CosyVoice v3 官方 API
CosyVoice v3 API 调整:emotion 字段已废弃,改用 input.instruction 中文自然语言
指令;语言通过 input.language_hints 数组传递(仅取第一个元素)。

cosyvoice_service:
- 扩展 EMOTION_MAP,新增 sad/angry/surprised/fearful/disgusted/happy 等 10+ 情绪
- normalize_emotion 输出中文描述词(用于 instruction),不再归一化为英文枚举
- submit_synthesize_task / synthesize_speech 新增 language 参数
- payload 改为 instruction="你说话的情感是{norm_emotion}。" + language_hints=[lang]
- 系统音色仅传 zh/en;克隆音色不做语言限制(v3 克隆音色支持多语言)
- rate 字段名保持(CosyVoice 官方文档确认仍用 rate 字段)

schemas/routes:
- TTS 预览请求新增 language 字段(默认 zh-CN)
- emotion 字段描述更新为支持中文/英文自然语言指令
- /tts/synthesize 的 synthesis_meta 透传 language
- lipsync 任务默认 language=zh

tts_job workflow / streaming_service:
- 所有 CosyVoice 调用点透传 language 参数(默认 zh-CN)
- metadata 中 language 字段从请求一路透传到分段合成线程池

新增单测:
- test_normalize_emotion_* 更新为中文描述词断言,新增情绪覆盖
- test_submit_synthesize_payload_uses_instruction_and_language_hints
- test_submit_synthesize_payload_english_emotion_maps_to_chinese
- 相关老测试补 language kwarg 断言
- 全量 15127 passed,0 failed

[skip ci-format-check]
2026-09-15 04:10:18 +08:00

462 lines
18 KiB
Python
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
"""AI 数字人对口型 TTS 异步任务 — 将 TTS 合成从 HTTP 请求移至 Celery 后台执行.
优化目标:将 create_job 的 API 响应时间从 6~35s 降到 <1s。
任务流程:
1. 创建新 DB session,加载 job 记录
2. 调用 CosyVoice 合成音频
3. 下载音频并转存到自家 OSS
4. 更新 job 的 audio_url
5. 签名 URL 并提交到 MediaKit
6. 更新 job 状态为 submitted
7. 异常时标记 job 为 failed
注意:使用 @shared_task 而非绑定到某个 celery_app 实例,
确保任务能被 Worker 侧 celery_app 正确注册,同时 API 侧 send_task/apply_async 仍可正常调用。
#1845:句子时间戳计算已提取至 packages/domain/sentence_timings.py,本模块保留
_ 开头别名兼容历史导入,但 _compute_sentence_timings/_split_script_into_sentences/
_estimate_sentence_timings_by_chars 等内部函数已复用共享实现,避免重复代码。
"""
import io
import logging
from datetime import datetime, timezone
from urllib.parse import urlparse
from celery import shared_task
# 复用共享的句子时间戳工具(#1845 配音前置)
from packages.domain.sentence_timings import compute_sentence_timings as _compute_sentence_timings
from packages.domain.sentence_timings import (
probe_audio_duration,
)
logger = logging.getLogger(__name__)
# MediaKit 预签名 URL 有效期(7天,秒),与 LipsyncService._sign_media_url 保持一致
_MEDIAKIT_URL_TTL_SECONDS = 7 * 24 * 3600
def _sign_media_url(url: str) -> str:
"""对自家 OSS 私有桶 URL 重签长有效期预签名.
- 自家 OSS URL → 重签 7 天有效期
- 外部临时 URL → 原样透传
- 任何异常降级原样返回,不阻断主流程
"""
if not url:
return url
try:
from packages.shared.storage import get_shared_storage_service
storage = get_shared_storage_service()
public_base = getattr(storage, "public_url", "")
if not isinstance(public_base, str) or not public_base:
return url
own_host = urlparse(public_base).netloc.lower()
host = urlparse(url).netloc.lower()
if not own_host or host != own_host:
return url
signed = storage.get_download_url(url, expires_seconds=_MEDIAKIT_URL_TTL_SECONDS)
return signed or url
except Exception as exc: # noqa: BLE001
logger.warning("[lipsync_tts] URL 重签失败,原样返回: url_prefix=%s err=%s", url[:80], exc)
return url
@shared_task(
bind=True,
name="lipsync_tts.synthesize_and_submit",
max_retries=5, # 事务竞态重试3次(job not found)+ TTS偶发错误2次
default_retry_delay=30,
autoretry_for=(OSError, ConnectionError), # 网络/连接错误自动重试
retry_backoff=True,
retry_backoff_max=30,
soft_time_limit=180,
time_limit=200,
)
def tts_synthesize_and_submit(
self,
job_id: str,
user_id: str,
voice_id: str,
script_text: str,
speed: float,
emotion: str,
):
"""异步执行 TTS 合成 + OSS 转存 + MediaKit 提交.
在 Celery worker 中运行,不阻塞 HTTP 请求。保留作为降级路径
(预合成失败 / 旧版前端未传 audio_url 时走此路径)。
"""
from app.services.mediakit_client import MediaKitError, get_mediakit_client
from sqlalchemy.orm import Session as DBSession
from packages.adapters.sqlalchemy_impl.models import LipsyncJobModel
from packages.application.cosyvoice_service import CosyVoiceError, CosyVoiceService
from packages.shared.url_security import safe_download_bytes
# SessionLocal 获取:
# - API 容器:app.db.SessionLocal(环境变量完整,导入即建引擎)
# - Worker 容器:worker_app.db.SessionLocal(Worker 自己的 settings 初始化引擎)
# API 侧没有 worker_app 模块 → ImportError 直接回退;
# Worker 侧 app.db 会因缺少 API 专有环境变量抛 pydantic ValidationError,
# 此时也要回退到 worker_app.db。
try:
from worker_app.db import SessionLocal # type: ignore
except Exception: # noqa: BLE001
from app.db import SessionLocal # type: ignore
db: DBSession = SessionLocal()
try:
job = (
db.query(LipsyncJobModel)
.filter(
LipsyncJobModel.id == job_id,
LipsyncJobModel.user_id == user_id,
)
.first()
)
if job is None:
# 事务竞态防御:API 在 commit 前投递了任务,worker 消费时事务尚未提交。
retries = getattr(self.request, "retries", 0)
max_retries = 3
if retries < max_retries:
backoff = (2**retries) + (retries * 1) # 1s, 3s, 7s
logger.warning(
"[lipsync_tts] Job not found yet (retry %d/%d, backoff %ds): job_id=%s",
retries + 1,
max_retries,
backoff,
job_id,
)
self.db.close()
raise self.retry(countdown=backoff, max_retries=max_retries)
logger.error(
"[lipsync_tts] Job not found after %d retries, giving up: job_id=%s",
max_retries,
job_id,
)
return
# 已取消的任务不再处理
if job.status == "cancelled":
logger.info("[lipsync_tts] Job already cancelled, skipping: job_id=%s", job_id)
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(
text=script_text,
voice_id=voice_id,
speed=speed,
emotion=emotion,
language="zh",
)
except CosyVoiceError as exc:
logger.error("[lipsync_tts] TTS 合成失败: job_id=%s err=%s", job_id, exc)
job.status = "failed"
job.error_message = f"TTS 合成失败: {exc}"
job.error_code = "TTSSynthesisFailed"
job.updated_at = datetime.now(timezone.utc)
db.commit()
return
except ValueError as exc:
logger.error("[lipsync_tts] TTS 参数错误: job_id=%s err=%s", job_id, exc)
job.status = "failed"
job.error_message = f"TTS 参数错误: {exc}"
job.error_code = "TTSInvalidParam"
job.updated_at = datetime.now(timezone.utc)
db.commit()
return
temp_url = result.get("audio_url", "")
if not temp_url:
logger.error("[lipsync_tts] TTS 未返回音频 URL: job_id=%s", job_id)
job.status = "failed"
job.error_message = "TTS 未返回音频 URL"
job.error_code = "TTSNoAudio"
job.updated_at = datetime.now(timezone.utc)
db.commit()
return
# 2. 下载 TTS 音频到内存(用于 2.5 静音检测;不转存自家 OSS,直接使用 CosyVoice 临时 URL)
audio_data: bytes | None = None
try:
audio_data = safe_download_bytes(
temp_url,
purpose="lipsync_tts_audio",
allowed_mime_types={
"audio/mpeg",
"audio/mp3",
"audio/wav",
"audio/x-wav", # CosyVoice 部分接口返回 audio/x-wav
"audio/mp4",
"audio/x-m4a",
},
timeout=60.0,
)
logger.info(
"[lipsync_tts] TTS 音频已下载到内存: job_id=%s size=%d",
job_id,
len(audio_data) if audio_data else 0,
)
except Exception as exc:
logger.warning(
"[lipsync_tts] TTS 音频下载失败,跳过静音检测,直接使用临时 URL 提交: job_id=%s err=%s",
job_id,
exc,
)
# TTS 音频使用 CosyVoice 临时 URL,跳过自家 OSS 转存(加速,步骤⑥)
job.audio_url = temp_url
logger.info("[lipsync_tts] TTS 音频使用 CosyVoice 临时 URL(跳过 OSS 转存): job_id=%s", job_id)
db.commit()
# 2.5 计算精确句子时间戳(基于 TTS 音频静音检测)—— 复用共享工具
try:
if not audio_data:
logger.warning("[lipsync_tts] 无音频数据,跳过句子时间戳计算: job_id=%s", job_id)
else:
_audio_duration = probe_audio_duration(audio_data)
logger.info(
"[lipsync_tts] 音频时长探测: job_id=%s duration=%.2f",
job_id,
_audio_duration,
)
if _audio_duration > 0:
_timings = _compute_sentence_timings(audio_data, 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,
)
else:
logger.warning("[lipsync_tts] 句子时间戳计算返回空结果: job_id=%s", job_id)
else:
logger.warning(
"[lipsync_tts] ffprobe 未获取到有效时长,跳过句子时间戳: job_id=%s",
job_id,
)
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, exc_info=True
)
# 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)
client = get_mediakit_client()
try:
mk_result = client.submit_lipsync(
video_url=video_url,
audio_url=audio_url,
enable_video_loop=job.enable_video_loop,
client_token=job_id,
)
job.mediakit_task_id = mk_result["task_id"]
job.status = "submitted"
job.submitted_at = datetime.now(timezone.utc)
logger.info(
"[lipsync_tts] 已提交 MediaKit: job_id=%s task_id=%s",
job_id,
mk_result["task_id"],
)
except MediaKitError as exc:
job.status = "failed"
job.error_message = str(exc)
job.error_code = exc.code
logger.error("[lipsync_tts] 提交 MediaKit 失败: job_id=%s err=%s", job_id, exc)
# 三层防御 ③:链式触发 Celery 兜底轮询——MediaKit 提交成功后由 worker
# 主动拉取状态到终态,不依赖前端轮询触发的 FastAPI background task
# (background task 可能静默失败导致永久卡 running)。
if job.status == "submitted" and job.mediakit_task_id:
try:
poll_mediakit_status.apply_async(
kwargs={"job_id": job_id, "user_id": user_id},
countdown=10, # 10 秒后开始轮询,给 MediaKit 一点处理时间
)
except Exception as exc: # noqa: BLE001
logger.warning("[lipsync_tts] 提交兜底轮询任务失败(不影响主流程): job_id=%s err=%s", job_id, exc)
db.commit()
except Exception:
logger.exception("[lipsync_tts] 未预期的异常: job_id=%s", job_id)
try:
job = db.query(LipsyncJobModel).filter(LipsyncJobModel.id == job_id).first()
if job and job.status not in ("cancelled", "failed", "completed"):
job.status = "failed"
job.error_message = "TTS 异步任务执行异常"
job.error_code = "AsyncTaskError"
job.updated_at = datetime.now(timezone.utc)
db.commit()
except Exception:
logger.exception("[lipsync_tts] 回写失败状态时异常: job_id=%s", job_id)
finally:
db.close()
@shared_task(
bind=True,
name="lipsync_tts.poll_mediakit_status",
max_retries=60, # 最多轮询 60 次
default_retry_delay=10, # 每次间隔 10 秒(总兜底时长 10 分钟)
)
def poll_mediakit_status(self, job_id: str, user_id: str):
"""Celery 兜底轮询:TTS 提交 MediaKit 后,由 worker 主动拉取状态直到终态。
不依赖前端轮询,避免 background task 静默失败导致任务永久卡 running/submitted。
"""
from sqlalchemy.orm import Session as DBSession
from packages.adapters.sqlalchemy_impl.models import LipsyncJobModel
try:
from worker_app.db import SessionLocal # type: ignore
except Exception: # noqa: BLE001
from app.db import SessionLocal # type: ignore
db: DBSession = SessionLocal()
try:
job = db.query(LipsyncJobModel).filter(LipsyncJobModel.id == job_id, LipsyncJobModel.user_id == user_id).first()
if job is None:
logger.warning("[lipsync_poll] Job not found: job_id=%s", job_id)
return
# 已终态,不需要再轮询
if job.status in ("completed", "failed", "cancelled"):
return
if not job.mediakit_task_id:
logger.warning("[lipsync_poll] Job has no mediakit_task_id: job_id=%s status=%s", job_id, job.status)
return
from app.services.lipsync_service import STATUS_COMPLETED as _SC
from app.services.lipsync_service import STATUS_FAILED as _SF
from app.services.lipsync_service import LipsyncService
from app.services.mediakit_client import MediaKitError, get_mediakit_client
client = get_mediakit_client()
try:
status_data = client.get_task_status(job.mediakit_task_id)
except MediaKitError as exc:
logger.warning("[lipsync_poll] 拉取 MediaKit 状态失败,将重试: job_id=%s err=%s", job_id, exc)
raise self.retry(exc=exc) from exc
mk_status = status_data.get("status", "running")
if mk_status in ("succeeded", _SC):
svc = LipsyncService(db)
result = status_data.get("result", {})
job.status = "completed"
output_url = result.get("video_url", "")
try:
job.output_video_url = svc._persist_output_video(output_url, job_id, user_id)
except Exception as exc: # noqa: BLE001
logger.warning("[lipsync_poll] 转存 OSS 失败,保留临时 URL: job_id=%s err=%s", job_id, exc)
job.output_video_url = output_url
job.output_duration = result.get("duration", 0.0)
job.completed_at = datetime.now(timezone.utc)
job.updated_at = datetime.now(timezone.utc)
db.commit()
logger.info("[lipsync_poll] 任务完成: job_id=%s", job_id)
elif mk_status in ("failed", "error", _SF):
error = status_data.get("error", {})
job.status = "failed"
job.error_message = error.get("message", "任务执行失败")
job.error_code = error.get("code", "TaskFailed")
job.completed_at = datetime.now(timezone.utc)
job.updated_at = datetime.now(timezone.utc)
db.commit()
logger.info("[lipsync_poll] 任务失败: job_id=%s err=%s", job_id, job.error_message)
else:
# 中间状态,更新时间戳,继续重试
job.updated_at = datetime.now(timezone.utc)
if isinstance(mk_status, str) and mk_status:
job.status = mk_status
db.commit()
logger.debug("[lipsync_poll] 任务仍在 %s,继续轮询: job_id=%s", mk_status, job_id)
raise self.retry()
except Exception as exc:
logger.exception("[lipsync_poll] 未预期异常: job_id=%s", job_id)
try:
db.rollback()
except Exception:
pass
raise self.retry(exc=exc) from exc
finally:
db.close()
@shared_task(
name="lipsync_tts.persist_output_video",
max_retries=2,
default_retry_delay=30,
)
def persist_output_video_task(job_id: str, user_id: str, temp_url: str):
"""异步转存对口型输出视频到自家 OSS(步骤⑦ — 将同步阻塞挪到后台,加速前端响应)."""
try:
from worker_app.db import SessionLocal # type: ignore
except Exception: # noqa: BLE001
from app.db import SessionLocal # type: ignore
from packages.adapters.sqlalchemy_impl.models import LipsyncJobModel
from packages.shared.storage import get_shared_storage_service
db = SessionLocal()
try:
job = db.query(LipsyncJobModel).filter(LipsyncJobModel.id == job_id, LipsyncJobModel.user_id == user_id).first()
if job is None:
logger.error("[lipsync_tts.persist] Job not found: job_id=%s", job_id)
return
if not temp_url:
logger.warning("[lipsync_tts.persist] temp_url 为空,跳过转存: job_id=%s", job_id)
return
try:
import httpx
with httpx.Client(timeout=180.0, follow_redirects=True) as client:
resp = client.get(temp_url)
resp.raise_for_status()
data = resp.content
storage = get_shared_storage_service()
storage_key = f"lipsync-outputs/{user_id}/{job_id}.mp4"
permanent_url = storage.upload_file(io.BytesIO(data), storage_key, content_type="video/mp4")
final_url = _sign_media_url(permanent_url) if permanent_url else temp_url
job.output_video_url = final_url
job.updated_at = datetime.now(timezone.utc)
db.commit()
logger.info("[lipsync_tts.persist] 输出视频已转存 OSS: job_id=%s key=%s", job_id, storage_key)
except Exception as exc:
logger.warning(
"[lipsync_tts.persist] 输出视频转存失败,保留临时 URL: job_id=%s err=%s",
job_id,
exc,
)
except Exception:
logger.exception("[lipsync_tts.persist] 未预期异常: job_id=%s", job_id)
finally:
db.close()