Skip to content

Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.

fix(cli): warn when jieba is unavailable for readable SRT splitting (#3726)

林SO940b0d246314
diff --git a/funasr/cli.py b/funasr/cli.py
index ecf7db9d4..cf9cad319 100644
--- a/funasr/cli.py
+++ b/funasr/cli.py
@@ -2,6 +2,7 @@
import argparse
import json
+import logging
import os
import re
import sys
@@ -184,13 +185,28 @@ def _timestamps_are_ordered(timestamps):
)
+_SUBTITLE_WORD_BOUNDARY_FALLBACK_WARNED = False
+
+
+def _warn_subtitle_word_boundary_fallback(error):
+ """Report once that subtitle splitting lost its word-boundary scoring."""
+ global _SUBTITLE_WORD_BOUNDARY_FALLBACK_WARNED
+ if _SUBTITLE_WORD_BOUNDARY_FALLBACK_WARNED:
+ return
+ _SUBTITLE_WORD_BOUNDARY_FALLBACK_WARNED = True
+ logging.warning(
+ f"jieba word segmentation is unavailable ({error}); readable subtitles "
+ "fall back to punctuation-only boundaries and can split words across "
+ "cues. Install jieba to restore word-aware subtitle splitting."
+ )
+
+
def _subtitle_break_weights(text, token_spans):
"""Map token-boundary indices to lexical and punctuation preferences."""
boundary_to_token = {span[1]: index + 1 for index, span in enumerate(token_spans)}
breaks = {len(token_spans): 0.0}
try:
- import logging
import warnings
with warnings.catch_warnings():
@@ -222,8 +238,8 @@ def _subtitle_break_weights(text, token_spans):
breaks[token_index] = max(
breaks.get(token_index, 0.0), strength
)
- except (ImportError, RuntimeError, ValueError):
- pass
+ except (ImportError, RuntimeError, ValueError) as error:
+ _warn_subtitle_word_boundary_fallback(error)
for index, span in enumerate(token_spans[:-1], 1):
boundary = span[1]
diff --git a/tests/test_cli.py b/tests/test_cli.py
index d67422851..f6fe42054 100644
--- a/tests/test_cli.py
+++ b/tests/test_cli.py
@@ -609,3 +609,20 @@ def test_merge_subtitle_segments_avoids_single_character_phrase_breaks():
for boundary in boundary_offsets
)
search_from = phrase_start + 1
+
+
+def test_subtitle_break_weights_warns_once_without_jieba(monkeypatch):
+ text = "钟书成为朴家保姆后他轻易的偷偷扭了一下儿子的耳朵"
+ token_spans = [[index, index + 1] for index in range(len(text))]
+ monkeypatch.setattr(cli, "_SUBTITLE_WORD_BOUNDARY_FALLBACK_WARNED", False, raising=False)
+
+ with (
+ patch.dict(sys.modules, {"jieba": None}),
+ patch("logging.warning") as warning,
+ ):
+ first = cli._subtitle_break_weights(text, token_spans)
+ second = cli._subtitle_break_weights(text, token_spans)
+
+ assert first == second == {len(token_spans): 0.0}
+ assert warning.call_count == 1
+ assert "jieba" in warning.call_args[0][0]