fix(voice): harden noisy streaming turns

2026-04-17 00:30:41 +05:00
parent 8ca5f23e93
commit 39096e114f
6 changed files with 267 additions and 15 deletions
@@ -24,6 +24,30 @@ def test_context_summary_tracks_city_and_branch_slot_for_schedule():
    assert summary["open_slots"] == ["branch_or_address"]


+def test_context_summary_maps_common_asr_city_mishearing_to_almaty():
+    summary = update_context_summary_from_user_turn(
+        None,
+        channel="voice",
+        language="ru",
+        customer_name="Ернур",
+        text="Хочу узнать график работы.",
+        now="2026-04-12T00:00:00+00:00",
+    )
+
+    updated = update_context_summary_from_user_turn(
+        json.dumps(summary, ensure_ascii=False),
+        channel="voice",
+        language="ru",
+        customer_name="Ернур",
+        text="Матта.",
+        now="2026-04-12T00:00:02+00:00",
+    )
+
+    assert updated["active_intent"] == "schedule"
+    assert updated["confirmed_facts"]["city"] == "Алмата"
+    assert updated["open_slots"] == ["branch_or_address"]
+
+
 def test_context_summary_does_not_overwrite_meaningful_request_with_low_signal():
    summary = update_context_summary_from_user_turn(
        None,
@@ -1655,6 +1655,44 @@ def test_voice_postprocess_reply_uses_summary_context_when_raw_window_lost_topic
    assert "филиал" in normalized or "адрес" in normalized


+def test_voice_decision_uses_summary_city_slot_instead_of_asking_city_again(monkeypatch):
+    def _unexpected_llm(messages):
+        raise AssertionError(f"LLM should not be called when summary slot prompt is enough: {messages!r}")
+
+    monkeypatch.setattr(ai_module, "_request_structured_model_decision", _unexpected_llm)
+
+    decision = voice_module._voice_decision(
+        language="ru",
+        customer=SimpleNamespace(display_name="Ернур"),
+        interaction=SimpleNamespace(interaction_id="int_voice_summary_slot", customer_id=None, status="open", queue_id=None, subject=None),
+        transcript_text="Матта.",
+        transcript_window=[
+            SimpleNamespace(speaker="caller", text="Матта.", sequence_no=1, source_type="voice_asr", barge_in_interrupted=False, created_at=utc_now_iso()),
+        ],
+        context_summary=json.dumps(
+            {
+                "customer_name": "Ернур",
+                "active_intent": "schedule",
+                "active_request_text": "Хочу узнать график работы",
+                "confirmed_facts": {"city": "Алмата", "branch_hint": None, "service_hint": "график работы", "request_number": None},
+                "open_slots": ["branch_or_address"],
+            },
+            ensure_ascii=False,
+        ),
+        kb_results=[],
+        disclosure_required=False,
+        customer_name_value="Ернур",
+        customer_name_status="name_obtained",
+        request_metadata={"voice_v2_enabled": True},
+    )
+
+    assert decision["intent"] == "clarification"
+    assert decision["needs_handoff"] is False
+    assert "Алмата" in decision["reply_text"]
+    assert "город" in decision["reply_text"].lower()
+    assert "филиал" in decision["reply_text"].lower() or "адрес" in decision["reply_text"].lower()
+
+
 def test_ai_enqueue_creates_outbound_ai_reply_and_delivery_flow(monkeypatch):
    monkeypatch.setenv("AI_TELEGRAM_ENABLED", "1")
    monkeypatch.setenv("AI_PROVIDER", "stub")
@@ -1047,6 +1047,12 @@ def test_media_runtime_voice_v2_emits_blind_ack_on_first_turn_without_partial_si
    assert speak_events[1][0] == "РџРѕРґСЃРєР°Р¶РёС‚Рµ РїРѕРґСЂРѕР±РЅРµРµ, РїРѕР¶Р°Р»СѓР№СЃС‚Р°."


+def test_media_runtime_low_signal_filter_catches_short_asr_noise():
+    assert AudioSocketMediaRuntime._is_low_signal_transcript("Давай")
+    assert AudioSocketMediaRuntime._is_low_signal_transcript("твой")
+    assert AudioSocketMediaRuntime._is_low_signal_transcript("Поргай, что это")
+
+
 def test_media_runtime_ignores_low_signal_utterance_without_ack_or_turn():
    planned: list[tuple[str, str, str, dict | None]] = []
    delivered: list[tuple[str, str, bool]] = []
@@ -1624,6 +1630,120 @@ def test_media_runtime_streaming_sidecar_push_does_not_block_vad_finalization():
    assert streaming_provider.push_count >= 1


+def test_media_runtime_streaming_partial_poll_does_not_block_turn_close_ack():
+    events: list[str] = []
+
+    class _StreamingProvider(StreamingASRProvider):
+        name = "streaming-sidecar"
+        supports_streaming = True
+
+        def __init__(self) -> None:
+            self.poll_count = 0
+
+        def poll_partial(self, stream_id: str) -> StreamingASRPartial | None:
+            del stream_id
+            self.poll_count += 1
+            time.sleep(0.25)
+            return StreamingASRPartial(text="need schedule", language="ru", confidence=0.8, is_stable=True)
+
+        def finalize(self, stream_id: str) -> ASRTranscription:
+            assert stream_id == "stream-1"
+            events.append("finalize")
+            return ASRTranscription(text="need schedule", language="ru", confidence=0.9)
+
+        def close_stream(self, stream_id: str) -> None:
+            assert stream_id == "stream-1"
+
+    streaming_provider = _StreamingProvider()
+    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(),
+        streaming_asr_provider=streaming_provider,
+        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="schedule",
+            reply_text="reply",
+            confidence=0.8,
+            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_reply(
+        actor: MediaActor,
+        text: str,
+        *,
+        is_greeting: bool,
+        style_hints: dict[str, object] | None = None,
+        reply_phase: str | None = "main",
+    ) -> None:
+        del actor, text, is_greeting, style_hints
+        events.append(str(reply_phase))
+
+    runtime._speak_reply = _fake_speak_reply  # type: ignore[method-assign]
+
+    async def _scenario() -> None:
+        actor = MediaActor(
+            registration=MediaRegistration(
+                voice_session_id="avs_media_runtime_turn_close_no_poll",
+                call_id="call_media_runtime_turn_close_no_poll",
+                interaction_id="int_media_runtime_turn_close_no_poll",
+                ai_session_id="ais_media_runtime_turn_close_no_poll",
+                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,
+                voice_v2_duplex=True,
+                voice_v2_streaming_asr_backend="local_sidecar",
+            ),
+            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.asr_streaming_enabled = True
+        actor.asr_stream_id = "stream-1"
+        pcm_frame = (1000).to_bytes(2, "little", signed=True) * 160
+        await runtime._process_utterance(actor, pcm_frame * 40, False)
+
+    asyncio.run(_scenario())
+
+    assert streaming_provider.poll_count == 0
+    assert events[:2] == ["ack", "finalize"]
+    assert "main" in events
+
+
 def test_media_runtime_merges_thinking_continuation_into_current_utterance():
    runtime = AudioSocketMediaRuntime(
        enabled=True,