Skip to content

FunASR-powered video transcription, subtitle generation, and LLM-assisted clipping tool with a local Gradio UI.

fix: clean SenseVoice tags and split long subtitles

游雁4a7ece922faf
diff --git a/funclip/videoclipper.py b/funclip/videoclipper.py
index b12c3a3..9a4cdb5 100644
--- a/funclip/videoclipper.py
+++ b/funclip/videoclipper.py
@@ -17,11 +17,16 @@ import moviepy.editor as mpy
from moviepy.video.tools.subtitles import SubtitlesClip, TextClip
from moviepy.editor import VideoFileClip, concatenate_videoclips
from moviepy.video.compositing.CompositeVideoClip import CompositeVideoClip
-from utils.subtitle_utils import generate_srt, generate_srt_clip
+from utils.subtitle_utils import generate_srt, generate_srt_clip, str2list
from utils.argparse_tools import ArgumentParser, get_commandline_args
from utils.trans_utils import pre_proc, proc, write_state, load_state, proc_spk, convert_pcm_to_float
+MAX_SUBTITLE_DURATION_MS = 8000
+MAX_SUBTITLE_TOKENS = 30
+SENSEVOICE_TAG_RE = re.compile(r"<\|[^|>]+\|>")
+
+
def _is_valid_timestamp(timestamp):
return (
isinstance(timestamp, list)
@@ -31,18 +36,75 @@ def _is_valid_timestamp(timestamp):
)
+def _clean_recognition_text(text):
+ if text is None:
+ return ""
+ text = SENSEVOICE_TAG_RE.sub("", str(text))
+ text = re.sub(r"\s+", " ", text).strip()
+ return text.strip("“”")
+
+
+def _split_long_sentence(sent):
+ timestamp = sent.get("timestamp")
+ if not _is_valid_timestamp(timestamp):
+ return []
+
+ cleaned_text = _clean_recognition_text(sent.get("text"))
+ normalized = dict(sent)
+ normalized["text"] = cleaned_text
+ normalized["timestamp"] = timestamp
+
+ tokens = str2list(cleaned_text)
+ if len(timestamp) <= 1 or len(tokens) != len(timestamp):
+ return [normalized]
+
+ chunks = []
+ start = 0
+ for idx in range(len(tokens)):
+ duration = timestamp[idx][1] - timestamp[start][0]
+ token_count = idx - start + 1
+ should_split = (
+ idx > start
+ and (
+ duration >= MAX_SUBTITLE_DURATION_MS
+ or token_count >= MAX_SUBTITLE_TOKENS
+ )
+ )
+ if should_split:
+ chunk = dict(normalized)
+ chunk["text"] = tokens[start : idx + 1]
+ chunk["timestamp"] = timestamp[start : idx + 1]
+ chunks.append(chunk)
+ start = idx + 1
+
+ if not chunks:
+ return [normalized]
+
+ if start < len(tokens):
+ chunk = dict(normalized)
+ chunk["text"] = tokens[start:]
+ chunk["timestamp"] = timestamp[start:]
+ chunks.append(chunk)
+
+ return chunks
+
+
def _normalize_recognition_result(result):
- text = result.get("text") or result.get("text_tn") or result.get("raw_text") or ""
- raw_text = result.get("raw_text") or result.get("text_tn") or text
- timestamp = result.get("timestamp") or []
+ text = _clean_recognition_text(
+ result.get("text") or result.get("text_tn") or result.get("raw_text") or ""
+ )
+ raw_text = _clean_recognition_text(
+ result.get("raw_text") or result.get("text_tn") or text
+ )
+ timestamp = result.get("timestamp") or result.get("timestamps") or []
sentence_info = []
for sent in result.get("sentence_info") or []:
if _is_valid_timestamp(sent.get("timestamp")):
- sentence_info.append(sent)
+ sentence_info.extend(_split_long_sentence(sent))
if not sentence_info and text and _is_valid_timestamp(timestamp):
- sentence_info = [{"text": text, "timestamp": timestamp}]
+ sentence_info = _split_long_sentence({"text": text, "timestamp": timestamp})
return text, raw_text, timestamp, sentence_info
diff --git a/tests/test_recognition_result_compat.py b/tests/test_recognition_result_compat.py
index 61c6e34..8880355 100644
--- a/tests/test_recognition_result_compat.py
+++ b/tests/test_recognition_result_compat.py
@@ -61,6 +61,43 @@ class TestRecognitionResultCompat(unittest.TestCase):
)
self.assertIn("测试文本", srt)
+ def test_sensevoice_rich_tags_are_removed_from_text_and_srt(self):
+ text, srt, state = self._recognize(
+ {
+ "key": "utt",
+ "text": "<|zh|><|NEUTRAL|><|Speech|><|woitn|>你好世界",
+ "timestamp": [[0, 300], [300, 600], [600, 900], [900, 1200]],
+ "sentence_info": [
+ {
+ "text": "<|zh|><|NEUTRAL|><|Speech|><|woitn|>你好世界",
+ "timestamp": [[0, 300], [300, 600], [600, 900], [900, 1200]],
+ }
+ ],
+ }
+ )
+
+ self.assertEqual(text, "你好世界")
+ self.assertEqual(state["sentences"][0]["text"], "你好世界")
+ self.assertNotIn("<|zh|>", state["recog_res_raw"])
+ self.assertNotIn("<|zh|>", srt)
+
+ def test_long_token_level_sentence_is_split_into_subtitle_chunks(self):
+ long_text = "一二三四五六七八九十甲乙"
+ timestamps = [[i * 1000, (i + 1) * 1000] for i in range(len(long_text))]
+
+ _, srt, state = self._recognize(
+ {
+ "key": "utt",
+ "text": long_text,
+ "timestamp": timestamps,
+ "sentence_info": [{"text": long_text, "timestamp": timestamps}],
+ }
+ )
+
+ self.assertGreater(len(state["sentences"]), 1)
+ self.assertIn("00:00:00,000 --> 00:00:08,000", srt)
+ self.assertIn("00:00:08,000 --> 00:00:12,000", srt)
+
if __name__ == "__main__":
unittest.main()