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): repack tiny-gap subtitle boundaries

LauraGPTdb959d301031
diff --git a/funasr/cli.py b/funasr/cli.py
index 6a58d27bd..d845d88b7 100644
--- a/funasr/cli.py
+++ b/funasr/cli.py
@@ -80,6 +80,7 @@ def merge_subtitle_segments(
"""Group sentence timestamps into bounded, readable subtitle cues."""
def can_follow(left, right):
left_text = str(left.get("text", ""))
+ right_text = str(right.get("text", ""))
gap_ms = right.get("start", 0) - left.get("end", 0)
left_speaker = left.get("speaker", left.get("spk"))
right_speaker = right.get("speaker", right.get("spk"))
@@ -89,6 +90,13 @@ def merge_subtitle_segments(
and (
left_text.rstrip().endswith(SUBTITLE_CONTINUATION_PUNCTUATION)
or _subtitle_body_length(left_text) <= 2
+ or (
+ gap_ms <= min(max_gap_ms, 100)
+ and _subtitle_body_length(right_text) > 2
+ and right_text.rstrip().endswith(
+ SUBTITLE_CONTINUATION_PUNCTUATION
+ )
+ )
)
)
diff --git a/tests/test_cli.py b/tests/test_cli.py
index db9ab38e8..dafbd25cf 100644
--- a/tests/test_cli.py
+++ b/tests/test_cli.py
@@ -238,6 +238,41 @@ def test_merge_subtitle_segments_keeps_continuation_chain_with_its_ending():
]
+def test_merge_subtitle_segments_repacks_false_sentence_boundary_after_tiny_gap():
+ segments = [
+ {"start": 148930, "end": 149830, "text": "就在八月二十五日,"},
+ {
+ "start": 150010,
+ "end": 151930,
+ "text": "上汽大通直接向平台发起投诉,",
+ },
+ {
+ "start": 152170,
+ "end": 153970,
+ "text": "投诉对象是那些转述基金会。",
+ },
+ {"start": 154030, "end": 155050, "text": "声明的自媒体文章,"},
+ {
+ "start": 155470,
+ "end": 157330,
+ "text": "上汽大通的投诉内容写的很清楚,",
+ },
+ ]
+
+ assert cli.merge_subtitle_segments(segments) == [
+ {
+ "start": 148930,
+ "end": 151930,
+ "text": "就在八月二十五日,上汽大通直接向平台发起投诉,",
+ },
+ {
+ "start": 152170,
+ "end": 157330,
+ "text": "投诉对象是那些转述基金会。声明的自媒体文章,上汽大通的投诉内容写的很清楚,",
+ },
+ ]
+
+
def test_merge_subtitle_segments_groups_two_character_sentence():
segments = [
{"start": 12450, "end": 12690, "text": "突然。"},