fix: 음성통화 STT 잡음/겹침 목소리 환각 완화

no_speech_prob가 높은 세그먼트를 텍스트 조합에서 제외하고
hallucination_silence_threshold를 걸어, 배경 소음이나 다른 사람
목소리가 섞였을 때 엉뚱한 텍스트가 연속으로 나오는 문제를 완화.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

src/tools/voice_engine.py

@@ -53,6 +53,11 @@ log = logging.getLogger('voice_engine')
 logging.getLogger('faster_whisper').setLevel(logging.WARNING)
 
 MIN_PARTIAL_BYTES = int(16000 * 2 * 0.6)  # ~0.6s of 16kHz mono PCM16 before we bother decoding
+# Segments where Whisper itself flags no_speech_prob above this are dropped before
+# joining the transcript — catches noise/background-voice hallucination (a garbled
+# guess is still emitted with low no_speech_prob when a second voice overlaps the
+# caller's, so this isn't a full fix, just a confidence floor).
+NO_SPEECH_PROB_CUTOFF = 0.6
 
 
 def is_unrecoverable_cuda_error(e: Exception) -> bool:
@@ -140,8 +145,9 @@ class Engine:
         try:
             segments, info = self.stt_model.transcribe(
                 path, language=language or None, beam_size=5, vad_filter=True, condition_on_previous_text=False,
+                hallucination_silence_threshold=2.0,
             )
-            text = ''.join(seg.text for seg in segments).strip()
+            text = ''.join(seg.text for seg in segments if seg.no_speech_prob < NO_SPEECH_PROB_CUTOFF).strip()
             return text, (info.language if info else language)
         finally:
             try:
@@ -173,8 +179,9 @@ class Engine:
         audio = pcm16_bytes_to_float(pcm)
         segments, info = self.stt_model.transcribe(
             audio, language=language or None, beam_size=1, vad_filter=True, condition_on_previous_text=False,
+            hallucination_silence_threshold=2.0,
         )
-        text = ''.join(seg.text for seg in segments).strip()
+        text = ''.join(seg.text for seg in segments if seg.no_speech_prob < NO_SPEECH_PROB_CUTOFF).strip()
         return text, (info.language if info else language)
 
     # ── "Streaming" TTS ─────────────────────────────────────────────────