Correlation updates

drb-c2-core/app/internal/transcription.py

@@ -121,12 +121,26 @@ def _sync_transcribe(
                 language="en",
                 prompt=prompt,
                 response_format="verbose_json",
+                temperature=0,
             )
         text = response.text.strip() or None
+
+        # Filter hallucinated segments.  Two sources of hallucination in P25 recordings:
+        #
+        # 1. Trailing silence / static — Whisper fills silence past real content with
+        #    sequential radio codes (10-4, 10-5...).  Clamped by audio duration.
+        #
+        # 2. Leading silence — OP25 recordings typically have a short silence at the
+        #    start before the first PTT press.  Whisper sometimes hallucinates filler
+        #    words or codes over this silence.  Detected via no_speech_prob > 0.8
+        #    (Whisper's own confidence that a segment contains no real speech).
+        audio_duration: float = getattr(response, "duration", None) or float("inf")
         segments = [
             {"start": round(s.start, 2), "end": round(s.end, 2), "text": s.text.strip()}
             for s in (response.segments or [])
             if s.text.strip()
+            and s.start < audio_duration
+            and getattr(s, "no_speech_prob", 0.0) < 0.8
         ]
         return text, segments
     finally: