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.
Python ai-video-generatorcontent-creationffmpeginstagram-reelsllmpythonshort-videosubtitlestext-to-speechtiktokvideo-automationvideo-workflowworkflow-automationyoutube-shorts
fix(voice): recognize pause tags with digits right after CJK keywords
Rudy Celeklid83b98af7133
diff --git a/app/utils/utils.py b/app/utils/utils.pyindex ae5066ff..5f148e9d 100644--- a/app/utils/utils.py+++ b/app/utils/utils.py@@ -321,9 +321,12 @@ PAUSE_TAG_KEYWORDS = (r"пауза|тишина|停顿|暂停|静音|ポーズ|一時停止|無音|일시중지|정지")# 匹配所有包含停顿关键词的标签(方括号或圆括号),无论其参数合法与否均匹配,-# 以确保非法标签(如 [pause: -2s]、[pause: nope]、[pause: 0s])在合成前被彻底清除而不会泄漏给 TTS+# 以确保非法标签(如 [pause: -2s]、[pause: nope]、[pause: 0s])在合成前被彻底清除而不会泄漏给 TTS。+# 关键词后不能用 `\b`:中日韩文字和数字都属于 `\w`,`[停顿3秒]` 这种不写冒号的+# 常见写法会因为没有词边界而漏匹配,标签原文被直接朗读。这里只禁止关键词后+# 紧跟字母(允许数字),`[paused]`、`[暂停一下]` 这类普通文字仍不会被误判。PAUSE_TAG_PATTERN = re.compile(- rf"[\[\(]\s*(?:{PAUSE_TAG_KEYWORDS})\b(?:\s*[::]?\s*([^\]\)]*?))?\s*[\]\)]",+ rf"[\[\(]\s*(?:{PAUSE_TAG_KEYWORDS})(?![^\W\d])(?:\s*[::]?\s*([^\]\)]*?))?\s*[\]\)]",re.IGNORECASE,)diff --git a/test/services/test_voice.py b/test/services/test_voice.pyindex 118b4e3e..8604cc87 100644--- a/test/services/test_voice.py+++ b/test/services/test_voice.py@@ -1899,6 +1899,29 @@ class TestElevenLabsVoice(unittest.TestCase):self.assertNotIn("(pausa", flex_cleaned)self.assertNotIn("[pause", flex_cleaned)+ def test_pause_tags_accept_duration_directly_after_cjk_keyword(self):+ """中日文习惯不写冒号或空格,`[停顿3秒]` 也必须识别为停顿而不是被朗读。"""+ script = "大家好。[停顿3秒]今天聊聊理财。[ポーズ2秒]最後。[pause2s]End."++ self.assertTrue(utils.has_pause_tags("[暂停2秒]"))+ self.assertEqual(+ utils.parse_script_with_pauses(script),+ [+ ("speech", "大家好。"),+ ("pause", 3.0),+ ("speech", "今天聊聊理财。"),+ ("pause", 2.0),+ ("speech", "最後。"),+ ("pause", 2.0),+ ("speech", "End."),+ ],+ )+ self.assertNotIn("停顿", utils.remove_pause_tags(script))++ # 关键词后紧跟字母的普通文字仍不是停顿标签。+ for text in ("[paused]", "(pauses)", "[暂停一下]"):+ self.assertFalse(utils.has_pause_tags(text), text)+def test_concat_audio_files_reads_pcm_in_bounded_chunks(self):"""Long pause-aware narration should not load each WAV into memory."""real_wave_open = vs.wave.open