fix(soniox): widen final span to all timed tokens, never regress end_time

livekit-plugins/livekit-plugins-soniox/livekit/plugins/soniox/stt.py

@@ -676,8 +676,10 @@ class _LangStats(NamedTuple):
 class _TokenAccumulator:
     """Accumulates token metadata (text, language, speaker, timing, confidence).
 
-    Tokens are assumed to arrive in chronological order, so start_time is taken
-    from the first token and end_time is continuously overwritten by the latest.
+    Tokens are assumed to arrive in chronological order, but individual tokens
+    may omit timing keys. start_time is the earliest ``start_ms`` seen and
+    end_time the latest ``end_ms`` seen, so a token that carries timing late
+    (or regresses) cannot collapse the span onto itself.
     """
 
     def __init__(self) -> None:
@@ -714,11 +716,13 @@ def update(self, token: dict[str, Any]) -> None:
             self.language = self._get_language()
         if "speaker" in token and self.speaker_id is None:
             self.speaker_id = str(token["speaker"])
-        if "start_ms" in token and not self._has_start_time:
-            self._has_start_time = True
-            self.start_time = float(token["start_ms"])
+        if "start_ms" in token:
+            start_ms = float(token["start_ms"])
+            if not self._has_start_time or start_ms < self.start_time:
+                self._has_start_time = True
+                self.start_time = start_ms
         if "end_ms" in token:
-            self.end_time = float(token["end_ms"])
+            self.end_time = max(self.end_time, float(token["end_ms"]))
         if "confidence" in token:
             self._confidence_sum += token["confidence"]
             self._confidence_count += 1

tests/test_plugin_soniox_stt.py

@@ -115,6 +115,67 @@ def test_token_accumulator_lang_segments_empty_initially():
     assert accumulator._lang_segments == []
 
 
+def test_token_accumulator_span_covers_all_timed_tokens():
+    """A token that carries timing late must not collapse the span onto itself.
+
+    Regression for #6885: when the leading tokens of an utterance arrive
+    without ``start_ms``/``end_ms`` and only a trailing token has them, the
+    emitted span used to collapse onto that trailing token (e.g. two words in
+    180ms).
+    """
+    from livekit.plugins.soniox.stt import _TokenAccumulator
+
+    accumulator = _TokenAccumulator()
+    accumulator.update({"text": " He's", "language": "en", "is_final": True})
+    accumulator.update({"text": " trying", "language": "en", "is_final": True})
+    accumulator.update(
+        {"text": ".", "language": "en", "is_final": True, "start_ms": 43663, "end_ms": 43843}
+    )
+
+    assert accumulator.text == " He's trying."
+    assert accumulator.start_time == 43663
+    assert accumulator.end_time == 43843
+
+
+def test_token_accumulator_end_time_never_regresses():
+    """A trailing token whose ``end_ms`` regresses must not pull the end back.
+
+    Regression for #6885: ``end_time`` was overwritten by every token that
+    carried ``end_ms``, so a non-monotonic trailing token shortened the span.
+    """
+    from livekit.plugins.soniox.stt import _TokenAccumulator
+
+    accumulator = _TokenAccumulator()
+    accumulator.update(
+        {"text": " He's", "language": "en", "is_final": True, "start_ms": 41000, "end_ms": 41400}
+    )
+    accumulator.update(
+        {"text": " trying", "language": "en", "is_final": True, "start_ms": 41400, "end_ms": 43843}
+    )
+    accumulator.update(
+        {"text": ".", "language": "en", "is_final": True, "start_ms": 43663, "end_ms": 43700}
+    )
+
+    assert accumulator.start_time == 41000
+    assert accumulator.end_time == 43843
+
+
+def test_token_accumulator_start_time_takes_earliest():
+    """The earliest ``start_ms`` wins even if a later token carries an earlier one."""
+    from livekit.plugins.soniox.stt import _TokenAccumulator
+
+    accumulator = _TokenAccumulator()
+    accumulator.update(
+        {"text": " He's", "language": "en", "is_final": True, "start_ms": 41000, "end_ms": 41400}
+    )
+    accumulator.update(
+        {"text": " trying", "language": "en", "is_final": True, "start_ms": 40000, "end_ms": 43843}
+    )
+
+    assert accumulator.start_time == 40000
+    assert accumulator.end_time == 43843
+
+
 # ---------------------------------------------------------------------------
 # _lang_segments_to_fields helper
 # ---------------------------------------------------------------------------