fix(stt): don't clear the primary speaker on an unattributed segment

livekit-agents/livekit/agents/stt/multi_speaker_adapter.py

@@ -283,10 +283,15 @@ def _get_rms_for_timerange(self, start_time: float, end_time: float) -> float |
         return float(np.median(self._rms_buffer[start:end]))
 
     def _update_primary_speaker(self, sd: SpeechData) -> None:
-        if sd.speaker_id is None or not self._detect_primary:
+        if not self._detect_primary:
             self._primary_speaker = None
             return
 
+        if sd.speaker_id is None:
+            # an unattributed segment carries no evidence about who is primary,
+            # so leave the current primary in place rather than clearing it
+            return
+
         rms = self._get_rms_for_timerange(sd.start_time, sd.end_time)
         if rms is None:
             return

tests/test_multi_speaker_primary_speaker.py

@@ -0,0 +1,88 @@
+import pytest
+
+from livekit import rtc
+from livekit.agents.stt import SpeechData
+from livekit.agents.stt.multi_speaker_adapter import (
+    PrimarySpeakerDetectionOptions,
+    _PrimarySpeakerDetector,
+)
+
+pytestmark = pytest.mark.unit
+
+SAMPLE_RATE = 1000
+SAMPLES_PER_FRAME = 100  # 100ms at 1000Hz, matches frame_size_ms below
+
+
+def _constant_frame(value: int) -> rtc.AudioFrame:
+    """Build a mono int16 frame where every sample equals ``value``."""
+    return rtc.AudioFrame(
+        data=value.to_bytes(2, "little", signed=True) * SAMPLES_PER_FRAME,
+        sample_rate=SAMPLE_RATE,
+        num_channels=1,
+        samples_per_channel=SAMPLES_PER_FRAME,
+    )
+
+
+def _speech_data(speaker_id: str | None, start_time: float, end_time: float) -> SpeechData:
+    return SpeechData(
+        language="en",
+        text="hello",
+        start_time=start_time,
+        end_time=end_time,
+        speaker_id=speaker_id,
+    )
+
+
+def _detector() -> _PrimarySpeakerDetector:
+    opts = PrimarySpeakerDetectionOptions(frame_size_ms=100, min_rms_samples=1)
+    return _PrimarySpeakerDetector(primary_detection_options=opts)
+
+
+def _push_speech(
+    detector: _PrimarySpeakerDetector,
+    speaker_id: str | None,
+    amplitude: int,
+    start_time: float,
+) -> None:
+    detector.push_audio(_constant_frame(amplitude))
+    detector.push_audio(_constant_frame(amplitude))
+    detector._update_primary_speaker(_speech_data(speaker_id, start_time, start_time + 0.2))
+
+
+class TestPrimarySpeakerReset:
+    def test_quiet_speaker_does_not_take_over(self) -> None:
+        """Control: a much quieter speaker never becomes primary."""
+        detector = _detector()
+
+        _push_speech(detector, "A", amplitude=3000, start_time=0.0)
+        assert detector._primary_speaker == "A"
+
+        _push_speech(detector, "B", amplitude=50, start_time=0.2)
+        assert detector._primary_speaker == "A"
+
+    def test_unattributed_segment_keeps_primary_speaker(self) -> None:
+        """A final transcript the diarizer could not attribute is not evidence
+        that the primary speaker stopped being primary.
+        """
+        detector = _detector()
+
+        _push_speech(detector, "A", amplitude=3000, start_time=0.0)
+        assert detector._primary_speaker == "A"
+
+        _push_speech(detector, None, amplitude=3000, start_time=0.2)
+        assert detector._primary_speaker == "A"
+
+        # Without the primary retained, B skips the loudness comparison entirely
+        # and is crowned as "first speaker" despite being 60x quieter.
+        _push_speech(detector, "B", amplitude=50, start_time=0.4)
+        assert detector._primary_speaker == "A"
+
+    def test_detect_primary_disabled_keeps_primary_unset(self) -> None:
+        opts = PrimarySpeakerDetectionOptions(frame_size_ms=100, min_rms_samples=1)
+        detector = _PrimarySpeakerDetector(
+            detect_primary_speaker=False, primary_detection_options=opts
+        )
+
+        _push_speech(detector, "A", amplitude=3000, start_time=0.0)
+
+        assert detector._primary_speaker is None