fix(soniox): enforce cue duration cap using word end timestamp

litellm/llms/soniox/common_utils.py

@@ -241,7 +241,8 @@ def _group_tokens_into_cues(
         subtitles never bridge pauses in speech,
       - adding the next word would exceed _CUE_MAX_CHARS of display width
         (~two subtitle lines; East-Asian wide characters count double), or
-      - the cue would span more than _CUE_MAX_DURATION_MS.
+      - adding the next word would make the cue span more than
+        _CUE_MAX_DURATION_MS.
     A cue also ends after sentence-final punctuation, which keeps cue breaks
     at natural seams. Cue timestamps come straight from token timestamps;
     words without timestamps stay attached to the surrounding cue.
@@ -276,8 +277,9 @@ def _group_tokens_into_cues(
             )
             gap_exceeded = start_ms is not None and cue_end is not None and (start_ms - cue_end) >= _CUE_GAP_MS
             chars_exceeded = _text_width(_cue_text(current)) + _text_width(word["text"]) > _CUE_MAX_CHARS
+            word_end = word["end_ms"] if word["end_ms"] is not None else start_ms
             duration_exceeded = (
-                start_ms is not None and cue_start is not None and (start_ms - cue_start) >= _CUE_MAX_DURATION_MS
+                word_end is not None and cue_start is not None and (word_end - cue_start) > _CUE_MAX_DURATION_MS
             )
             if speaker_changed or gap_exceeded or chars_exceeded or duration_exceeded:
                 _flush()

tests/test_litellm/llms/soniox/audio_transcription/test_soniox_audio_transcription_transformation.py

@@ -455,6 +455,19 @@ class TestCueGroupingAlignment:
         assert "Guten" not in result
         assert "00:00:00,000 --> 00:00:00,600" in result
 
+    def test_should_split_before_word_whose_end_crosses_duration_cap(self):
+        from litellm.llms.soniox.common_utils import render_soniox_tokens_as_srt
+
+        tokens = [{"text": " hm", "start_ms": i * 650, "end_ms": i * 650 + 600} for i in range(10)] + [
+            {"text": " boom", "start_ms": 6900, "end_ms": 7600}
+        ]
+        result = render_soniox_tokens_as_srt(tokens)
+        cues = result.strip().split("\n\n")
+        assert len(cues) == 2
+        assert "00:00:00,000 --> 00:00:06,450" in cues[0]
+        assert "00:00:06,900 --> 00:00:07,600" in cues[1]
+        assert cues[1].endswith("boom")
+
     def test_should_keep_untimestamped_word_in_cue(self):
         from litellm.llms.soniox.common_utils import render_soniox_tokens_as_srt