fix(audio): detect transcript-wide hallucinations

tests/test_speech_quality.py

@@ -3,6 +3,7 @@ from __future__ import annotations
 from types import SimpleNamespace
 
 from imagefind.speech_quality import (
+    aggregate_transcript_quality,
     normalize_language,
     pcm16_voiced_regions,
     select_language,
@@ -10,6 +11,32 @@ from imagefind.speech_quality import (
 )
 
 
+def test_aggregate_quality_rejects_repeated_short_hallucinations() -> None:
+    segments = ([{"text": "好"}] * 11) + ([{"text": "拜拜"}] * 7) + ([{"text": "正常对话内容"}] * 4)
+    result = aggregate_transcript_quality(segments, base_score=1.0)
+    assert result.score < 0.7
+    assert "whole_repeated_phrase" in result.flags
+    assert "whole_short_segment_dominance" in result.flags
+    assert "whole_ending_hallucination" in result.flags
+    assert ("拜拜", 7) in result.repeated_phrases
+
+
+def test_sports_samples_patterns_cannot_remain_ready() -> None:
+    sports_2 = ([{"text": "好"}] * 11) + ([{"text": "拜拜"}] * 7) + ([{"text": "嗯"}] * 5)
+    sports_2 += [{"text": f"正常对话{i}"} for i in range(19)]
+    sports_3 = ([{"text": "拜拜"}] * 8) + ([{"text": "啊"}] * 7) + ([{"text": "嗯"}] * 7)
+    sports_3 += ([{"text": "谢谢大家收看"}] * 3) + [{"text": f"正常内容{i}"} for i in range(32)]
+    for segments in (sports_2, sports_3):
+        result = aggregate_transcript_quality(segments, base_score=1.0)
+        assert result.score < 0.7
+        assert "whole_repeated_phrase" in result.flags
+
+
+def test_traditional_ending_phrase_is_flagged() -> None:
+    result = aggregate_transcript_quality([{"text": "謝謝大家收看"}] * 3, base_score=1.0)
+    assert "whole_ending_hallucination" in result.flags
+
+
 def test_language_policy_treats_stream_metadata_as_weak_hint() -> None:
     assert select_language(["en", "zh"], policy="zh_priority") == "zh"
     assert select_language(["en", "zh"], policy="auto") == "en"