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

tests/test_ai_voice_media_runtime.py

@@ -1249,6 +1249,288 @@ def test_media_runtime_voice_v2_emits_blind_ack_on_first_turn_without_partial_si
     assert speak_events[1][0] == "Подскажите подробнее, пожалуйста."
 
 
+def test_media_runtime_voice_v2_skips_filler_ack_when_caller_says_goodbye():
+    planned: list[tuple[str, str, str, dict | None]] = []
+
+    class _GoodbyeASRProvider(ASRProvider):
+        name = "goodbye-asr"
+
+        def transcribe(self, audio_bytes: bytes, *, language_hint: str | None = None) -> ASRTranscription:
+            assert audio_bytes
+            return ASRTranscription(text="Спасибо, до свидания", language=language_hint or "ru", confidence=0.9)
+
+    runtime = AudioSocketMediaRuntime(
+        enabled=True,
+        host="127.0.0.1",
+        port=0,
+        frame_ms=20,
+        idle_timeout_seconds=2.0,
+        registration_wait_timeout_seconds=0.5,
+        min_speech_ms=40,
+        trailing_silence_ms=40,
+        max_turn_ms=400,
+        asr_provider=_GoodbyeASRProvider(),
+        tts_provider=_StubTTSProvider(),
+        load_registration_by_media_uuid=lambda value: None,
+        mark_media_connected=lambda session_id, value: None,
+        mark_media_ended=lambda session_id, reason: None,
+        touch_media_frame=lambda session_id: None,
+        set_state=lambda session_id, state, handoff_reason, metadata: None,
+        get_pending_greeting=lambda session_id: None,
+        mark_reply_delivered=lambda session_id, text, is_greeting: None,
+        plan_reply=lambda session_id, text, metadata, kind: planned.append((session_id, text, kind, metadata)),
+        process_turn=lambda session_id, transcript_text, language, barge_in, metadata: (
+            time.sleep(0.25)
+            or VoiceAITurnDecisionOut(
+                language=language or "ru",
+                intent="closing",
+                reply_text="Хорошо, всего доброго!",
+                confidence=0.9,
+                needs_handoff=False,
+                handoff_reason=None,
+                case_action="close",
+                kb_refs=[],
+                summary_text="call wrapped up",
+                model="stub-voice",
+                latency_ms=1,
+                status="active",
+            )
+        ),
+        request_handoff=lambda session_id, customer_request_text, decision: None,
+        handle_media_error=lambda session_id, message, metadata: None,
+    )
+
+    async def _fake_speak_text(
+        current_actor,
+        text: str,
+        *,
+        is_greeting: bool,
+        style_hints: dict[str, object] | None = None,
+    ) -> None:
+        del current_actor, is_greeting, style_hints
+        await asyncio.sleep(0)
+
+    runtime._speak_text = _fake_speak_text  # type: ignore[method-assign]
+    pcm_frame = (1000).to_bytes(2, "little", signed=True) * 160
+
+    async def _scenario() -> None:
+        actor = MediaActor(
+            registration=MediaRegistration(
+                voice_session_id="avs_media_runtime_v2_goodbye",
+                call_id="call_media_runtime_v2_goodbye",
+                interaction_id="int_media_runtime_v2_goodbye",
+                ai_session_id="ais_media_runtime_v2_goodbye",
+                language="ru",
+                media_uuid=str(uuid.uuid4()),
+                queue_code="voice_lab_ai",
+                queue_id="que_voice_lab_ai",
+                agent_profile="voice_support",
+                voice_v2_enabled=True,
+                voice_v2_ack_mode="immediate_short",
+                voice_v2_streaming_tts=True,
+                voice_v2_partial_asr=False,
+            ),
+            reader=asyncio.StreamReader(),
+            writer=None,  # type: ignore[arg-type]
+            vad=EnergyVAD(frame_ms=20, min_speech_ms=40, trailing_silence_ms=40, max_turn_ms=400),
+            frame_ms=20,
+            frame_bytes=320,
+        )
+        actor.finalized_caller_turn_count = 1
+        await runtime._process_utterance(actor, pcm_frame, False)
+
+    asyncio.run(_scenario())
+
+    assert [item[2] for item in planned] == ["reply"]
+    assert planned[0][1] == "Хорошо, всего доброго!"
+
+
+def test_media_runtime_voice_v2_throttles_repeated_filler_ack_within_gap():
+    planned: list[tuple[str, str, str, dict | None]] = []
+
+    class _SlowASRProvider(ASRProvider):
+        name = "slow-asr"
+
+        def transcribe(self, audio_bytes: bytes, *, language_hint: str | None = None) -> ASRTranscription:
+            assert audio_bytes
+            return ASRTranscription(text="Хочу узнать график работы", language=language_hint or "ru", confidence=0.9)
+
+    runtime = AudioSocketMediaRuntime(
+        enabled=True,
+        host="127.0.0.1",
+        port=0,
+        frame_ms=20,
+        idle_timeout_seconds=2.0,
+        registration_wait_timeout_seconds=0.5,
+        min_speech_ms=40,
+        trailing_silence_ms=40,
+        max_turn_ms=400,
+        asr_provider=_SlowASRProvider(),
+        tts_provider=_StubTTSProvider(),
+        load_registration_by_media_uuid=lambda value: None,
+        mark_media_connected=lambda session_id, value: None,
+        mark_media_ended=lambda session_id, reason: None,
+        touch_media_frame=lambda session_id: None,
+        set_state=lambda session_id, state, handoff_reason, metadata: None,
+        get_pending_greeting=lambda session_id: None,
+        mark_reply_delivered=lambda session_id, text, is_greeting: None,
+        plan_reply=lambda session_id, text, metadata, kind: planned.append((session_id, text, kind, metadata)),
+        process_turn=lambda session_id, transcript_text, language, barge_in, metadata: (
+            time.sleep(0.25)
+            or VoiceAITurnDecisionOut(
+                language=language or "ru",
+                intent="clarification",
+                reply_text="Подскажите, пожалуйста, какой город вас интересует?",
+                confidence=0.9,
+                needs_handoff=False,
+                handoff_reason=None,
+                case_action="keep_open",
+                kb_refs=[],
+                summary_text="reply ready",
+                model="stub-voice",
+                latency_ms=1,
+                status="active",
+            )
+        ),
+        request_handoff=lambda session_id, customer_request_text, decision: None,
+        handle_media_error=lambda session_id, message, metadata: None,
+    )
+
+    async def _fake_speak_text(
+        current_actor,
+        text: str,
+        *,
+        is_greeting: bool,
+        style_hints: dict[str, object] | None = None,
+    ) -> None:
+        del current_actor, is_greeting, style_hints
+        await asyncio.sleep(0)
+
+    runtime._speak_text = _fake_speak_text  # type: ignore[method-assign]
+    pcm_frame = (1000).to_bytes(2, "little", signed=True) * 160
+
+    async def _scenario() -> None:
+        actor = MediaActor(
+            registration=MediaRegistration(
+                voice_session_id="avs_media_runtime_v2_throttle",
+                call_id="call_media_runtime_v2_throttle",
+                interaction_id="int_media_runtime_v2_throttle",
+                ai_session_id="ais_media_runtime_v2_throttle",
+                language="ru",
+                media_uuid=str(uuid.uuid4()),
+                queue_code="voice_lab_ai",
+                queue_id="que_voice_lab_ai",
+                agent_profile="voice_support",
+                voice_v2_enabled=True,
+                voice_v2_ack_mode="immediate_short",
+                voice_v2_streaming_tts=True,
+                voice_v2_partial_asr=False,
+            ),
+            reader=asyncio.StreamReader(),
+            writer=None,  # type: ignore[arg-type]
+            vad=EnergyVAD(frame_ms=20, min_speech_ms=40, trailing_silence_ms=40, max_turn_ms=400),
+            frame_ms=20,
+            frame_bytes=320,
+        )
+        actor.finalized_caller_turn_count = 1
+        await runtime._process_utterance(actor, pcm_frame, False)
+        runtime._reset_live_turn_state(actor)
+        await runtime._process_utterance(actor, pcm_frame, False)
+
+    asyncio.run(_scenario())
+
+    assert [item[2] for item in planned] == ["ack", "reply", "reply"]
+
+
+def test_media_runtime_voice_v2_blind_ack_defers_to_known_low_signal_partial_transcript():
+    speak_events: list[str] = []
+
+    runtime = AudioSocketMediaRuntime(
+        enabled=True,
+        host="127.0.0.1",
+        port=0,
+        frame_ms=20,
+        idle_timeout_seconds=2.0,
+        registration_wait_timeout_seconds=0.5,
+        min_speech_ms=40,
+        trailing_silence_ms=40,
+        max_turn_ms=2000,
+        asr_provider=_StubASRProvider(),
+        tts_provider=_StubTTSProvider(),
+        load_registration_by_media_uuid=lambda value: None,
+        mark_media_connected=lambda session_id, value: None,
+        mark_media_ended=lambda session_id, reason: None,
+        touch_media_frame=lambda session_id: None,
+        set_state=lambda session_id, state, handoff_reason, metadata: None,
+        get_pending_greeting=lambda session_id: None,
+        mark_reply_delivered=lambda session_id, text, is_greeting: None,
+        plan_reply=lambda session_id, text, metadata, kind: None,
+        process_turn=lambda session_id, transcript_text, language, barge_in, metadata: VoiceAITurnDecisionOut(
+            language=language or "ru",
+            intent="clarification",
+            reply_text="Подскажите подробнее, пожалуйста.",
+            confidence=0.9,
+            needs_handoff=False,
+            handoff_reason=None,
+            case_action="keep_open",
+            kb_refs=[],
+            summary_text="reply ready",
+            model="stub-voice",
+            latency_ms=1,
+            status="active",
+        ),
+        request_handoff=lambda session_id, customer_request_text, decision: None,
+        handle_media_error=lambda session_id, message, metadata: None,
+    )
+
+    async def _fake_speak_text(
+        current_actor,
+        text: str,
+        *,
+        is_greeting: bool,
+        style_hints: dict[str, object] | None = None,
+    ) -> None:
+        del current_actor, is_greeting, style_hints
+        speak_events.append(text)
+        await asyncio.sleep(0)
+
+    runtime._speak_text = _fake_speak_text  # type: ignore[method-assign]
+    pcm_frame = (1000).to_bytes(2, "little", signed=True) * 160
+
+    async def _scenario() -> None:
+        actor = MediaActor(
+            registration=MediaRegistration(
+                voice_session_id="avs_media_runtime_v2_low_signal_blind",
+                call_id="call_media_runtime_v2_low_signal_blind",
+                interaction_id="int_media_runtime_v2_low_signal_blind",
+                ai_session_id="ais_media_runtime_v2_low_signal_blind",
+                language="ru",
+                media_uuid=str(uuid.uuid4()),
+                queue_code="voice_lab_ai",
+                queue_id="que_voice_lab_ai",
+                agent_profile="voice_support",
+                voice_v2_enabled=True,
+                voice_v2_ack_mode="immediate_short",
+                voice_v2_streaming_tts=True,
+                voice_v2_partial_asr=True,
+            ),
+            reader=asyncio.StreamReader(),
+            writer=None,  # type: ignore[arg-type]
+            vad=EnergyVAD(frame_ms=20, min_speech_ms=40, trailing_silence_ms=40, max_turn_ms=2000),
+            frame_ms=20,
+            frame_bytes=320,
+        )
+        actor.finalized_caller_turn_count = 1
+        # A caller who already said a recognized filler-answer ("да") should not
+        # get a blind ack just because the audio clip crossed the length threshold.
+        actor.stable_partial_transcript = "да"
+        await runtime._process_utterance(actor, pcm_frame * 40, False)
+
+    asyncio.run(_scenario())
+
+    assert speak_events == ["Подскажите подробнее, пожалуйста."]
+
+
 def test_media_runtime_low_signal_filter_catches_short_asr_noise():
     assert AudioSocketMediaRuntime._is_low_signal_partial_transcript("Давай")
     assert AudioSocketMediaRuntime._is_low_signal_partial_transcript("твой")