fix(voice): race slow streaming asr finalize

services/ai_orchestrator_service/voice.py

@@ -607,8 +607,12 @@ def _voice_reply_with_name(language: str, reply_text: str, name: str | None) ->
     short_name = _voice_short_name(name)
     if not short_name:
         return reply_text
+    reply_text = _voice_strip_reply_greeting_prefix(reply_text, name=short_name)
     prefix = _voice_disclosure_prefix(language)
     normalized_short = _voice_text_key(short_name)
+    first_words = _voice_text_key(" ".join(str(reply_text or "").split()[:8])).split()
+    if normalized_short in first_words:
+        return reply_text
     if reply_text.startswith(prefix):
         rest = reply_text[len(prefix) :].lstrip()
         if _voice_text_key(rest).startswith(normalized_short):
@@ -1344,6 +1348,28 @@ def _voice_is_midcall_greeting_reply(text: str | None) -> bool:
     )
 
 
+def _voice_strip_reply_greeting_prefix(reply_text: str | None, *, name: str | None = None) -> str:
+    text = str(reply_text or "").strip()
+    if not text:
+        return text
+    greeting_pattern = (
+        r"^\s*(?:здравствуйте|здравствуй|привет(?:ствую)?|"
+        r"добрый\s+(?:день|вечер|утро)|сәлеметсіз\s+бе|сәлем)\b[\s,!.:-]*"
+    )
+    stripped = re.sub(greeting_pattern, "", text, count=1, flags=re.IGNORECASE).strip()
+    if name and stripped != text:
+        short_name = _voice_short_name(name)
+        if short_name:
+            stripped = re.sub(
+                rf"^\s*{re.escape(short_name)}\b[\s,!.:-]*",
+                "",
+                stripped,
+                count=1,
+                flags=re.IGNORECASE,
+            ).strip()
+    return stripped or text
+
+
 def _voice_contains_false_lookup_promise(text: str | None) -> bool:
     normalized = _voice_text_key(text)
     if not normalized:
@@ -1388,15 +1414,16 @@ def _voice_postprocess_reply_text(
     prior_caller_texts = _voice_caller_context_before_current(caller_texts, transcript_text)
     active_topic_texts = _voice_service_context_texts(prior_caller_texts or caller_texts)
     active_topic_prompt = _voice_topic_prompt(language, active_topic_texts) if active_topic_texts else None
+    has_assistant_context = any(segment.speaker == "assistant" for segment in transcript_window)
     if _voice_is_hearing_check_caller_text(transcript_text):
         return _voice_hearing_check_reply(language, prior_caller_texts or caller_texts)
     if _voice_contains_false_lookup_promise(normalized_reply) and not kb_results:
         return active_topic_prompt or _voice_topic_prompt(language, caller_texts) or _voice_generic_prompt(language)
-    if _voice_is_midcall_greeting_reply(normalized_reply) and any(
-        segment.speaker == "assistant" for segment in transcript_window
-    ):
+    if _voice_is_midcall_greeting_reply(normalized_reply) and has_assistant_context:
         context_texts = prior_caller_texts if _voice_is_low_signal_caller_text(transcript_text) else caller_texts
         return _voice_topic_prompt(language, context_texts) or _voice_generic_prompt(language)
+    if has_assistant_context or _voice_has_service_topic(transcript_text):
+        normalized_reply = _voice_strip_reply_greeting_prefix(normalized_reply)
     if (
         active_topic_prompt
         and not _voice_has_service_topic(transcript_text)