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): preserve punctuation inside Whisper words

Rudy Celekliaee3197ced1a
diff --git a/app/services/subtitle.py b/app/services/subtitle.py
index 3bde812d..3f3419ee 100644
--- a/app/services/subtitle.py
+++ b/app/services/subtitle.py
@@ -12,6 +12,7 @@ except ImportError:
from loguru import logger
from app.config import config
+from app.models import const
from app.utils import utils
model_size = config.whisper.get("model_size", "large-v3")
@@ -132,15 +133,13 @@ def create(
is_segmented = True
seg_end = word.end
- # If it contains punctuation, then break the sentence.
+ # Accumulate words; only trailing punctuation ends a sentence.
seg_text += word.word
- if utils.str_contains_punctuation(word.word):
- # remove last char
- seg_text = seg_text[:-1]
- if not seg_text:
- continue
-
+ if word.word.rstrip().endswith(tuple(const.PUNCTUATIONS)):
+ # Punctuation inside a word (3.14, 1,000, 12:30) is not a
+ # sentence boundary. Remove only actual trailing delimiters.
+ seg_text = seg_text.rstrip().rstrip("".join(const.PUNCTUATIONS))
recognized(seg_text, seg_start, seg_end)
is_segmented = False
diff --git a/test/services/test_subtitle.py b/test/services/test_subtitle.py
index 3aeda5a6..09771abf 100644
--- a/test/services/test_subtitle.py
+++ b/test/services/test_subtitle.py
@@ -179,6 +179,56 @@ class TestSubtitleService(unittest.TestCase):
self.assertEqual([item[2] for item in items], ["Hello world", "Again"])
+ def test_create_preserves_numeric_punctuation_and_sentence_timing(self):
+ words = [
+ SimpleNamespace(start=0.0, end=0.2, word="Value"),
+ SimpleNamespace(start=0.2, end=0.5, word=" 3.14"),
+ SimpleNamespace(start=0.5, end=0.8, word=" costs"),
+ SimpleNamespace(start=0.8, end=1.1, word=" 1,000"),
+ SimpleNamespace(start=1.1, end=1.4, word=" dollars."),
+ SimpleNamespace(start=1.5, end=1.8, word="Next!"),
+ ]
+ fake_model = SimpleNamespace(
+ transcribe=lambda *_args, **_kwargs: (
+ [SimpleNamespace(start=0.0, end=1.8, words=words)],
+ SimpleNamespace(language="en", language_probability=0.99),
+ )
+ )
+ with tempfile.TemporaryDirectory() as tmp_dir:
+ output = Path(tmp_dir) / "numeric.srt"
+ with patch.object(subtitle, "model", fake_model), patch.object(
+ subtitle, "WhisperModel", object()
+ ):
+ subtitle.create("audio.mp3", str(output))
+ items = subtitle.file_to_subtitles(str(output))
+ self.assertEqual(
+ items,
+ [
+ (1, "00:00:00,000 --> 00:00:01,400", "Value 3.14 costs 1,000 dollars"),
+ (2, "00:00:01,500 --> 00:00:01,800", "Next"),
+ ],
+ )
+
+ def test_create_removes_only_trailing_sentence_punctuation(self):
+ words = [
+ SimpleNamespace(start=0.0, end=0.3, word="3.14?!"),
+ SimpleNamespace(start=0.4, end=0.7, word="1,000"),
+ ]
+ fake_model = SimpleNamespace(
+ transcribe=lambda *_args, **_kwargs: (
+ [SimpleNamespace(start=0.0, end=0.7, words=words)],
+ SimpleNamespace(language="en", language_probability=0.99),
+ )
+ )
+ with tempfile.TemporaryDirectory() as tmp_dir:
+ output = Path(tmp_dir) / "suffix.srt"
+ with patch.object(subtitle, "model", fake_model), patch.object(
+ subtitle, "WhisperModel", object()
+ ):
+ subtitle.create("audio.mp3", str(output))
+ items = subtitle.file_to_subtitles(str(output))
+ self.assertEqual([item[2] for item in items], ["3.14", "1,000"])
+
def test_create_word_level_writes_each_whisper_word_with_its_timing(self):
"""逐词模式应保留 Whisper 的每个词及其独立起止时间。"""
transcribe_kwargs = {}