feat(voice): emit immediate ack on turn close

2026-04-11 01:40:04 +05:00
parent c39fad7e2d
commit 6413beeba1
2 changed files with 165 additions and 50 deletions
@@ -722,3 +722,95 @@ def test_media_runtime_voice_v2_uses_partial_asr_to_start_ack_before_full_asr():
    assert speak_events
    assert speak_events[0][0] == runtime._ack_text("ru", "understanding")
    assert speak_events[0][1] < timings["full_finished"]
+
+
+def test_media_runtime_voice_v2_emits_generic_ack_before_full_asr_without_partial_signal():
+    timings: dict[str, float] = {}
+    speak_events: list[tuple[str, float]] = []
+
+    class _SlowOnlyASRProvider(ASRProvider):
+        name = "slow-only-asr"
+
+        def transcribe(self, audio_bytes: bytes, *, language_hint: str | None = None) -> ASRTranscription:
+            assert audio_bytes
+            timings["full_started"] = time.monotonic()
+            time.sleep(0.35)
+            timings["full_finished"] = time.monotonic()
+            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=_SlowOnlyASRProvider(),
+        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) -> None:
+        del current_actor, is_greeting
+        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_generic_ack",
+                call_id="call_media_runtime_v2_generic_ack",
+                interaction_id="int_media_runtime_v2_generic_ack",
+                ai_session_id="ais_media_runtime_v2_generic_ack",
+                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 * 30, False)
+
+    asyncio.run(_scenario())
+
+    assert speak_events
+    assert speak_events[0][0] == runtime._ack_text("ru", "generic")
+    assert speak_events[0][1] < timings["full_finished"]