STT bugfix

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

@@ -113,6 +113,10 @@ def _sync_transcribe(
         tg_prefix    = f"Talkgroup: {talkgroup_name}. " if talkgroup_name else ""
         prompt       = tg_prefix + vocab_prefix + _WHISPER_PROMPT
 
+        # Only whisper-1 supports verbose_json (per-segment timestamps + no_speech_prob).
+        # Newer models (gpt-4o-transcribe, gpt-4o-mini-transcribe) only accept json/text.
+        use_verbose = settings.stt_model == "whisper-1"
+
         openai_client = OpenAI(api_key=settings.openai_api_key)
         with open(tmp_path, "rb") as f:
             response = openai_client.audio.transcriptions.create(
@@ -120,9 +124,11 @@ def _sync_transcribe(
                 file=f,
                 language="en",
                 prompt=prompt,
-                response_format="verbose_json",
+                response_format="verbose_json" if use_verbose else "json",
                 temperature=0,
             )
+
+        if use_verbose:
             # Filter hallucinated segments.  Two sources of hallucination in P25 recordings:
             #
             # 1. Trailing silence / static — Whisper fills silence past real content with
@@ -141,12 +147,15 @@ def _sync_transcribe(
                 and getattr(s, "no_speech_prob", 0.0) < 0.8
             ]
             # Reconstruct text from non-hallucinated segments only so the two stay
-        # in sync.  If every segment was filtered (e.g. pure static or repeated
+            # in sync.  If every segment was filtered, text becomes None which prevents
-        # prompt-word hallucination like "Standby. Standby. Standby..."), text
+            # the intelligence pipeline from running on hallucinated content.
-        # becomes None which prevents the intelligence pipeline from running on
-        # hallucinated content.
             text = " ".join(s["text"] for s in segments) or None
             return text, segments
+        else:
+            # json format returns just {"text": "..."} — no segments or timestamps.
+            # Intelligence extraction falls back to treating the whole transcript as one block.
+            text = (response.text or "").strip() or None
+            return text, []
     finally:
         try:
             os.unlink(tmp_path)