fix(voice): stabilize v2 streaming topic handling

2026-04-12 21:59:32 +05:00
parent 7f4c80add1
commit 989bf9a0c1
6 changed files with 254 additions and 8 deletions
@@ -1497,6 +1497,45 @@ def test_voice_postprocess_reply_rewrites_false_lookup_promise():
    assert "филиал" in reply_text.lower() or "адрес" in reply_text.lower()


+def test_voice_postprocess_reply_rewrites_midcall_greeting_with_followup_question():
+    reply_text = voice_module._voice_postprocess_reply_text(
+        language="ru",
+        transcript_text="Здравствуйте, мне надо узнать...",
+        transcript_window=[
+            SimpleNamespace(speaker="assistant", text=voice_module._voice_greeting("ru"), sequence_no=1, source_type="voice_policy", barge_in_interrupted=False, created_at=utc_now_iso()),
+            SimpleNamespace(speaker="caller", text="Здравствуйте, мне надо узнать...", sequence_no=2, source_type="voice_asr", barge_in_interrupted=False, created_at=utc_now_iso()),
+        ],
+        reply_text="Здравствуйте, Ернур! О чем именно вы хотите узнать?",
+        kb_results=[],
+        needs_handoff=False,
+    )
+
+    normalized = reply_text.lower()
+    assert "здравствуйте" not in normalized
+    assert "о чем именно" not in normalized
+
+
+def test_voice_postprocess_reply_reuses_active_topic_after_frustration_turn():
+    reply_text = voice_module._voice_postprocess_reply_text(
+        language="ru",
+        transcript_text="Сколько раз повторять тебе?",
+        transcript_window=[
+            SimpleNamespace(speaker="caller", text="Мне надо узнать график работы.", sequence_no=1, source_type="voice_asr", barge_in_interrupted=False, created_at=utc_now_iso()),
+            SimpleNamespace(speaker="caller", text="В городе Алма-Ата.", sequence_no=2, source_type="voice_asr", barge_in_interrupted=False, created_at=utc_now_iso()),
+            SimpleNamespace(speaker="caller", text="Меня интересует филиал в Ауэзовском районе.", sequence_no=3, source_type="voice_asr", barge_in_interrupted=False, created_at=utc_now_iso()),
+            SimpleNamespace(speaker="caller", text="Сколько раз повторять тебе?", sequence_no=4, source_type="voice_asr", barge_in_interrupted=False, created_at=utc_now_iso()),
+        ],
+        reply_text="Пожалуйста, уточните, какая услуга вас интересует, чтобы я мог подсказать тариф.",
+        kb_results=[],
+        needs_handoff=False,
+    )
+
+    normalized = reply_text.lower()
+    assert "тариф" not in normalized
+    assert "услуг" not in normalized
+    assert "филиал" in normalized or "адрес" in normalized
+
+
 def test_turn_voice_session_early_plan_does_not_persist_partial_turns(monkeypatch):
    monkeypatch.setenv("AI_VOICE_POLICY_MODE", "v2_streaming_duplex")

@@ -624,6 +624,7 @@ def test_media_runtime_voice_v2_plays_short_ack_before_main_reply():
            frame_ms=20,
            frame_bytes=320,
        )
+        actor.finalized_caller_turn_count = 1
        await runtime._process_utterance(actor, pcm_frame, False)

    asyncio.run(_scenario())
@@ -842,6 +843,7 @@ def test_media_runtime_voice_v2_emits_generic_ack_before_full_asr_without_partia
            frame_ms=20,
            frame_bytes=320,
        )
+        actor.finalized_caller_turn_count = 1
        await runtime._process_utterance(actor, pcm_frame * 45, False)

    asyncio.run(_scenario())
@@ -940,6 +942,7 @@ def test_media_runtime_voice_v2_emits_ack_for_short_utterance_after_reduced_thre
            frame_ms=20,
            frame_bytes=320,
        )
+        actor.finalized_caller_turn_count = 1
        await runtime._process_utterance(actor, pcm_frame * 40, False)

    asyncio.run(_scenario())
@@ -949,6 +952,100 @@ def test_media_runtime_voice_v2_emits_ack_for_short_utterance_after_reduced_thre
    assert speak_events[0][1] < timings["full_finished"]


+def test_media_runtime_voice_v2_skips_blind_ack_on_first_turn_without_partial_signal():
+    speak_events: list[tuple[str, float]] = []
+
+    class _FastASRProvider(ASRProvider):
+        name = "fast-asr"
+
+        def transcribe(self, audio_bytes: bytes, *, language_hint: str | None = None) -> ASRTranscription:
+            assert audio_bytes
+            time.sleep(0.08)
+            return ASRTranscription(text="hello there", 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=2000,
+        asr_provider=_FastASRProvider(),
+        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="РџРѕРґСЃРєР°Р¶РёС‚Рµ РїРѕРґСЂРѕР±РЅРµРµ, РїРѕР¶Р°Р»СѓР№СЃС‚Р°.",
+            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, time.monotonic()))
+        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_first_turn",
+                call_id="call_media_runtime_v2_first_turn",
+                interaction_id="int_media_runtime_v2_first_turn",
+                ai_session_id="ais_media_runtime_v2_first_turn",
+                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=2000),
+            frame_ms=20,
+            frame_bytes=320,
+        )
+        await runtime._process_utterance(actor, pcm_frame * 40, False)
+
+    asyncio.run(_scenario())
+
+    assert len(speak_events) == 1
+    assert speak_events[0][0] == "РџРѕРґСЃРєР°Р¶РёС‚Рµ РїРѕРґСЂРѕР±РЅРµРµ, РїРѕР¶Р°Р»СѓР№СЃС‚Р°."
+
+
 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]] = []
@@ -1120,6 +1217,7 @@ def test_media_runtime_voice_v2_inserts_small_gap_between_ack_and_main_reply():
            frame_ms=20,
            frame_bytes=320,
        )
+        actor.finalized_caller_turn_count = 1
        await runtime._process_utterance(actor, pcm_frame * 40, False)

    asyncio.run(_scenario())
@@ -1226,6 +1324,7 @@ def test_media_runtime_voice_v2_emotive_ack_uses_ru_variants_and_style_hints_onl
            frame_ms=20,
            frame_bytes=320,
        )
+        actor.finalized_caller_turn_count = 1
        pcm_frame = (1000).to_bytes(2, "little", signed=True) * 160
        await runtime._process_utterance(actor, pcm_frame, False)

@@ -133,6 +133,49 @@ def test_streaming_asr_sidecar_stream_lifecycle(monkeypatch):
    assert len(engine.calls) >= 3


+def test_streaming_asr_sidecar_partial_uses_recent_audio_tail(monkeypatch):
+    engine = _reset_sidecar(monkeypatch)
+    monkeypatch.setattr(sidecar_module, "_partial_max_audio_ms", lambda: 1800)
+    client = TestClient(sidecar_module.app)
+
+    open_response = client.post(
+        "/internal/asr/streams",
+        headers=_admin_headers(),
+        json={
+            "session_id": "avs_sidecar_tail",
+            "language_hint": "ru",
+            "sample_rate_hz": 8000,
+            "encoding": "pcm_s16le",
+        },
+    )
+    assert open_response.status_code == 200
+    stream_id = open_response.json()["stream_id"]
+
+    long_chunk = _pcm_chunk(3200)
+    push_response = client.post(
+        f"/internal/asr/streams/{stream_id}/chunks",
+        headers=_admin_headers(),
+        json={
+            "pcm_b64": base64.b64encode(long_chunk).decode("ascii"),
+            "sample_rate_hz": 8000,
+            "encoding": "pcm_s16le",
+        },
+    )
+    assert push_response.status_code == 200
+
+    partial_response = client.get(
+        f"/internal/asr/streams/{stream_id}/partial",
+        headers=_admin_headers(),
+    )
+    assert partial_response.status_code == 200
+    assert engine.calls
+    partial_len, sample_rate_hz, language_hint = engine.calls[-1]
+    assert sample_rate_hz == 8000
+    assert language_hint == "ru"
+    assert partial_len < len(long_chunk)
+    assert partial_len <= int(1.8 * 8000 * 2)
+
+
 def test_streaming_asr_sidecar_rejects_unsupported_language(monkeypatch):
    _reset_sidecar(monkeypatch, supported_languages={"ru"})
    client = TestClient(sidecar_module.app)