Skip to content

利用 AI 大模型和自动化工作流,根据主题或关键词一键生成高清短视频。Generate HD short videos from a topic or keyword with an automated AI workflow.

fix(subtitle): prevent unexpected Whisper fallback

harry07031d5fa132480e
diff --git a/app/services/task.py b/app/services/task.py
index 7ed4ea39..210b1909 100644
--- a/app/services/task.py
+++ b/app/services/task.py
@@ -203,6 +203,10 @@ def generate_subtitle(task_id, params, video_script, sub_maker, audio_file):
subtitle_provider = config.app.get("subtitle_provider", "edge").strip().lower()
logger.info(f"\n\n## generating subtitle, provider: {subtitle_provider}")
+ if not subtitle_provider:
+ logger.info("subtitle provider is empty, skip subtitle generation")
+ return ""
+
if sub_maker is None and subtitle_provider != "whisper":
# 自定义音频不会经过 TTS,因此没有 Edge/Azure 等 TTS 返回的
# sub_maker 时间轴。只有 Whisper 可以直接从音频文件转写字幕;
@@ -213,16 +217,22 @@ def generate_subtitle(task_id, params, video_script, sub_maker, audio_file):
)
return ""
- subtitle_fallback = False
if subtitle_provider == "edge":
voice.create_subtitle(
text=video_script, sub_maker=sub_maker, subtitle_file=subtitle_path
)
if not os.path.exists(subtitle_path):
- subtitle_fallback = True
- logger.warning("subtitle file not found, fallback to whisper")
+ # Edge 字幕偶尔会因为时间轴与文案无法匹配而没有产出文件。这里不能
+ # 自动切换到 Whisper,否则首次失败会在用户不知情的情况下下载数 GB
+ # 的模型。只有显式配置 Whisper 时才允许加载模型,Edge 失败则保留
+ # 无字幕视频并记录原因,避免意外的网络和磁盘开销。
+ logger.warning(
+ "edge subtitle generation did not produce a subtitle file; "
+ "skip subtitles without falling back to whisper"
+ )
+ return ""
- if subtitle_provider == "whisper" or subtitle_fallback:
+ if subtitle_provider == "whisper":
subtitle.create(audio_file=audio_file, subtitle_file=subtitle_path)
logger.info("\n\n## correcting subtitle")
subtitle.correct(subtitle_file=subtitle_path, video_script=video_script)
diff --git a/config.example.toml b/config.example.toml
index 59f6a963..763fbe40 100644
--- a/config.example.toml
+++ b/config.example.toml
@@ -209,7 +209,9 @@ pollinations_model_name = ""
mimo_tts_model_name = "mimo-v2.5-tts"
mimo_tts_style_prompt = "请用自然、清晰、适合短视频旁白的语气朗读。"
-# Subtitle provider: "edge" or "whisper". Leave empty to skip generation.
+# Subtitle provider: "edge" or "whisper". Whisper is loaded only when selected
+# explicitly; Edge failures do not trigger an automatic Whisper model download.
+# Leave empty to skip generation.
subtitle_provider = "edge"
# FFmpeg is normally downloaded and detected automatically. If detection fails,
diff --git a/test/services/test_task.py b/test/services/test_task.py
index e8c25ed2..af477d4a 100644
--- a/test/services/test_task.py
+++ b/test/services/test_task.py
@@ -470,6 +470,48 @@ class TestTaskService(unittest.TestCase):
create_subtitle.assert_not_called()
whisper_create.assert_not_called()
+ def test_generate_subtitle_does_not_fallback_to_whisper_when_edge_fails(self):
+ """
+ Edge 没有生成字幕文件时应保留无字幕结果,不能自动下载 Whisper 模型。
+
+ 该场景可能由 TTS 时间轴与原始文案无法匹配触发。自动回退会让未选择
+ Whisper 的用户意外下载数 GB 模型,因此必须验证 Whisper 完全不会被调用。
+ """
+ task_id = "test-edge-subtitle-without-output"
+ task_dir = utils.task_dir(task_id)
+ params = VideoParams(
+ video_subject="edge subtitle",
+ video_script="Hello world.",
+ subtitle_enabled=True,
+ )
+ sub_maker = object()
+
+ try:
+ with (
+ patch.object(
+ tm.config,
+ "app",
+ dict(tm.config.app, subtitle_provider="edge"),
+ ),
+ patch.object(tm.voice, "create_subtitle") as create_subtitle,
+ patch.object(tm.subtitle, "create") as whisper_create,
+ patch.object(tm.subtitle, "correct") as whisper_correct,
+ ):
+ subtitle_path = tm.generate_subtitle(
+ task_id=task_id,
+ params=params,
+ video_script="Hello world.",
+ sub_maker=sub_maker,
+ audio_file=os.path.join(task_dir, "audio.mp3"),
+ )
+ finally:
+ shutil.rmtree(task_dir, ignore_errors=True)
+
+ self.assertEqual(subtitle_path, "")
+ create_subtitle.assert_called_once()
+ whisper_create.assert_not_called()
+ whisper_correct.assert_not_called()
+
def test_start_returns_each_intermediate_result(self):
"""
API 的 script、terms、audio、subtitle 和 materials 模式共用同一条任务