feat: enhance AudioSocketMediaRuntime to skip filler acks for closing intents and throttle repeated filler acks

services/ai_voice_runtime_service/media_runtime.py

@@ -217,6 +217,10 @@ class AudioSocketMediaRuntime:
         self._immediate_ack_min_ms = 700
         self._v2_ack_post_gap_seconds = 0.10
         self._v1_ack_wait_seconds = 0.6
+        # Skip a would-be filler if the previous one finished too recently, so rapid
+        # back-and-forth turns (e.g. a caller spelling out a phone number field by
+        # field) don't get a filler read before every single fragment.
+        self._ack_min_repeat_gap_seconds = 2.5
         self._partial_poll_interval_seconds = 0.20
         # Small startup cushion for streamed TTS playback: absorb ElevenLabs
         # network delivery jitter before we start pacing frames out to the
@@ -314,10 +318,32 @@ class AudioSocketMediaRuntime:
             return True
         return normalized in cls._FINAL_LOW_SIGNAL_PHRASES
 
+    _CLOSING_INTENT_PHRASES = (
+        "до свидания",
+        "всего доброго",
+        "хорошего дня",
+        "хорошего вечера",
+        "прощайте",
+        "созвонимся",
+        "это все спасибо",
+        "это всё спасибо",
+        "у меня все спасибо",
+        "у меня всё спасибо",
+        "больше вопросов нет",
+        "вопросов больше нет",
+        "спасибо за помощь",
+        "спасибо большое до свидания",
+        "сау болыңыз",
+        "келесіге дейін",
+        "рахмет көп",
+    )
+
     def _detect_early_intent(self, text: str) -> str:
         normalized = self._normalize_intent_text(text)
         if not normalized:
             return "unknown"
+        if any(token in normalized for token in self._CLOSING_INTENT_PHRASES):
+            return "closing"
         if any(token in normalized for token in ("оператор", "оператором", "человеком", "менеджер", "сотрудник")):
             return "operator_request"
         if any(token in normalized for token in ("график", "распис", "время работы", "work schedule", "жұмыс")):
@@ -347,6 +373,8 @@ class AudioSocketMediaRuntime:
 
     @staticmethod
     def _ack_kind_for_intent(intent: str) -> str:
+        if intent == "closing":
+            return "closing"
         if intent == "operator_request":
             return "handoff"
         if intent in {"schedule", "address", "price", "status", "problem"}:
@@ -406,8 +434,18 @@ class AudioSocketMediaRuntime:
             return True
         return normalized_intent != "unknown"
 
-    def _should_emit_blind_ack(self, actor: MediaActor, pcm_bytes: bytes) -> bool:
+    def _should_emit_blind_ack(self, actor: MediaActor, pcm_bytes: bytes, partial_transcript: str) -> bool:
+        """Duration-only fallback for when no usable partial transcript exists yet.
+
+        Must defer to the transcript when one *is* available: otherwise a caller
+        who already said a recognized filler-answer ("да"/"нет"/"хорошо") still
+        gets a blind ack just because the audio happened to cross the length
+        threshold, even though `_should_emit_partial_ack` correctly said no.
+        """
         del actor
+        transcript_text = str(partial_transcript or "").strip()
+        if transcript_text and self._is_low_signal_partial_transcript(transcript_text):
+            return False
         return len(pcm_bytes) >= self._immediate_ack_min_bytes
 
     @staticmethod
@@ -1116,10 +1154,22 @@ class AudioSocketMediaRuntime:
         metadata: dict[str, Any],
         ack_source: str,
         ack_kind: str | None = None,
+        intent: str | None = None,
     ) -> None:
         if actor.closed or actor.early_ack_started:
             return
-        ack_kind = ack_kind or self._ack_kind_for_intent(actor.partial_intent or "unknown")
+        effective_intent = str(
+            intent or actor.stable_partial_intent or actor.partial_intent or "unknown"
+        ).strip() or "unknown"
+        if effective_intent == "closing":
+            # The caller is wrapping up; a "thinking" filler right before the
+            # closing reply reads as robotic, so skip it and go straight to the reply.
+            return
+        if actor.last_ack_completed_monotonic and (
+            time.monotonic() - actor.last_ack_completed_monotonic
+        ) < self._ack_min_repeat_gap_seconds:
+            return
+        ack_kind = ack_kind or self._ack_kind_for_intent(effective_intent)
         ack_text, style_hints, ack_variant = self._select_ack_payload(
             actor,
             language=language,
@@ -1620,8 +1670,9 @@ class AudioSocketMediaRuntime:
                     language=actor.registration.language,
                     metadata=base_metadata,
                     ack_source="streaming_partial" if actor.asr_streaming_enabled else "precomputed_partial_asr",
+                    intent=partial_intent,
                 )
-            elif self._should_emit_blind_ack(actor, pcm_bytes):
+            elif self._should_emit_blind_ack(actor, pcm_bytes, partial_transcript):
                 await self._emit_early_ack(
                     actor,
                     language=actor.registration.language,