feat(voice): add elevenlabs stt and transcript truth fixes

2026-04-18 18:25:07 +05:00
parent ac2269b6a4
commit 34b807e460
9 changed files with 807 additions and 18 deletions
@@ -3104,7 +3104,8 @@ def test_custom_voice_name_texts_are_used_in_start_and_followup():
            metadata={"voice_start_language": "ru"},
        ),
    )
-    assert "Как к вам обращаться сейчас?" in decision.reply_text
+    assert "как к вам обращаться сейчас" not in decision.reply_text.lower()
+    assert "город" in decision.reply_text.lower()


 def test_downstream_voice_start_personalizes_greeting_and_finalizes_confirmed_name():
@@ -3176,7 +3177,8 @@ def test_downstream_voice_turn_adds_inline_name_followup_then_finalizes_provided
    )

    assert first_turn.metadata["customer_name_status"] == "name_not_obtained"
-    assert "как мне к вам обращаться" in first_turn.reply_text.lower()
+    assert "как мне к вам обращаться" not in first_turn.reply_text.lower()
+    assert "город" in first_turn.reply_text.lower()

    second_turn = voice_module.turn_voice_session(
        seeded["session_id"],
@@ -3212,6 +3214,34 @@ def test_downstream_voice_turn_adds_inline_name_followup_then_finalizes_provided
        session.close()


+def test_downstream_voice_turn_extracts_explicit_name_without_restarting_name_flow():
+    seeded = seed_voice_downstream_session(
+        marker=f"voice_explicit_name_intent_{new_id('seed')}",
+        name_status="name_not_obtained",
+        customer_display_name="+77010009999",
+    )
+
+    turn = voice_module.turn_voice_session(
+        seeded["session_id"],
+        VoiceAITurnIn(
+            voice_session_id=seeded["session_id"],
+            call_id=seeded["call_id"],
+            interaction_id=seeded["interaction_id"],
+            transcript_text="Меня зовут Ания, мне нужно узнать график работы",
+            language="ru",
+            sequence_no=1,
+            metadata={"voice_start_language": "ru"},
+        ),
+    )
+
+    assert turn.metadata["customer_name_status"] == "name_obtained"
+    assert turn.metadata["customer_name_value"] == "Ания"
+    assert "город" in turn.reply_text.lower()
+    assert "филиал" in turn.reply_text.lower()
+    assert "какой вопрос по работе" not in turn.reply_text.lower()
+    assert "как мне к вам обращаться" not in turn.reply_text.lower()
+
+
@pytest.mark.parametrize(
    ("transcript_text", "expected_name"),
    [
@@ -2,6 +2,9 @@ from __future__ import annotations

 import time

+import httpx
+import pytest
+
 from services.ai_voice_runtime_service.audiosocket import pcm16le_to_wav_bytes
 from services.ai_voice_runtime_service.providers import asr as asr_module

@@ -78,6 +81,99 @@ def test_yandex_asr_provider_posts_lpcm_with_tts_credential_fallback(monkeypatch
    assert calls[0]["timeout"] == 5.0


+def test_elevenlabs_asr_provider_posts_wav_with_tts_credential_fallback(monkeypatch):
+    calls: list[dict] = []
+    pcm = b"\x01\x00" * 160
+    wav_bytes = pcm16le_to_wav_bytes(pcm, sample_rate_hz=8000)
+
+    class _DummyClient:
+        def __init__(self, *, timeout: float) -> None:
+            self.timeout = timeout
+
+        def __enter__(self) -> "_DummyClient":
+            return self
+
+        def __exit__(self, exc_type, exc, tb) -> None:
+            return None
+
+        def post(
+            self,
+            url: str,
+            *,
+            headers: dict[str, str],
+            data: dict[str, str],
+            files: dict[str, tuple[str, bytes, str]],
+        ) -> _DummyResponse:
+            calls.append(
+                {
+                    "url": url,
+                    "headers": headers,
+                    "data": data,
+                    "files": files,
+                    "timeout": self.timeout,
+                }
+            )
+            return _DummyResponse({"text": "schedule", "language_code": "rus"})
+
+    monkeypatch.delenv("AI_VOICE_ASR_ELEVENLABS_API_KEY", raising=False)
+    monkeypatch.setenv("AI_VOICE_TTS_ELEVENLABS_API_KEY", "tts-elevenlabs-key")
+    monkeypatch.setenv("AI_VOICE_ASR_ELEVENLABS_API_BASE", "https://api.elevenlabs.example")
+    monkeypatch.setenv("AI_VOICE_ASR_ELEVENLABS_MODEL_ID", "scribe_v1")
+    monkeypatch.setattr(asr_module.httpx, "Client", _DummyClient)
+
+    provider = asr_module.ElevenLabsASRProvider()
+    result = provider.transcribe(wav_bytes, language_hint="ru")
+
+    assert result.text == "schedule"
+    assert result.language == "ru"
+    assert len(calls) == 1
+    assert calls[0]["url"] == "https://api.elevenlabs.example/v1/speech-to-text"
+    assert calls[0]["headers"]["xi-api-key"] == "tts-elevenlabs-key"
+    assert calls[0]["data"]["model_id"] == "scribe_v1"
+    assert calls[0]["data"]["language_code"] == "rus"
+    assert calls[0]["files"]["file"][0] == "turn.wav"
+    assert calls[0]["files"]["file"][2] == "audio/wav"
+    assert calls[0]["timeout"] == 20.0
+
+
+def test_elevenlabs_asr_provider_surfaces_http_auth_errors(monkeypatch):
+    request = httpx.Request("POST", "https://api.elevenlabs.example/v1/speech-to-text")
+    response = httpx.Response(401, request=request)
+
+    class _DummyClient:
+        def __init__(self, *, timeout: float) -> None:
+            self.timeout = timeout
+
+        def __enter__(self) -> "_DummyClient":
+            return self
+
+        def __exit__(self, exc_type, exc, tb) -> None:
+            return None
+
+        def post(
+            self,
+            url: str,
+            *,
+            headers: dict[str, str],
+            data: dict[str, str],
+            files: dict[str, tuple[str, bytes, str]],
+        ) -> _DummyResponse:
+            del url, headers, data, files
+            class _FailingResponse(_DummyResponse):
+                def raise_for_status(self_nonlocal) -> None:
+                    raise httpx.HTTPStatusError("401 Unauthorized", request=request, response=response)
+
+            return _FailingResponse()
+
+    monkeypatch.setenv("AI_VOICE_ASR_ELEVENLABS_API_KEY", "asr-elevenlabs-key")
+    monkeypatch.setattr(asr_module.httpx, "Client", _DummyClient)
+
+    provider = asr_module.ElevenLabsASRProvider()
+
+    with pytest.raises(httpx.HTTPStatusError):
+        provider.transcribe(b"\x01\x00" * 160, language_hint="ru")
+
+
 def test_yandex_asr_provider_uses_iam_token(monkeypatch):
    calls: list[dict] = []

@@ -325,6 +421,7 @@ def test_yandex_grpc_streaming_provider_returns_partial_and_final():


 def test_yandex_asr_builders():
+    assert isinstance(asr_module.build_asr_provider("elevenlabs"), asr_module.ElevenLabsASRProvider)
    assert isinstance(asr_module.build_asr_provider("yandex"), asr_module.YandexSpeechKitASRProvider)
    assert isinstance(asr_module.build_asr_provider("speechkit"), asr_module.YandexSpeechKitASRProvider)
    assert isinstance(
@@ -1082,7 +1082,7 @@ def test_voice_ai_summary_exposes_customer_name_state_for_operator(tmp_path):
                    confidence=0.96,
                    is_final=True,
                    barge_in_interrupted=False,
-                    payload_json="{}",
+                    payload_json='{"intent_bearing": true, "provider": "elevenlabs"}',
                    created_at=now,
                ),
                VoiceTranscriptSegmentRow(
@@ -1097,7 +1097,7 @@ def test_voice_ai_summary_exposes_customer_name_state_for_operator(tmp_path):
                    confidence=None,
                    is_final=True,
                    barge_in_interrupted=False,
-                    payload_json="{}",
+                    payload_json='{"delivery_state": "delivered", "provider": "elevenlabs"}',
                    created_at=now,
                ),
            ]
@@ -1113,6 +1113,173 @@ def test_voice_ai_summary_exposes_customer_name_state_for_operator(tmp_path):
    assert summary.customer_name_value == "Айдос"
    assert summary.customer_name_source == "voice_start"
    assert summary.voice_start_language == "ru"
+    assert summary.customer_request_text == "Соедините с оператором"
+    assert summary.ai_outcome_text == "Сейчас переведу вас на оператора."
+    assert [segment.text for segment in summary.transcript_segments] == [
+        "Соедините с оператором",
+        "Сейчас переведу вас на оператора.",
+    ]
+
+
+def test_voice_ai_summary_uses_intent_bearing_caller_turns_and_delivered_assistant_segments(tmp_path):
+    now = utc_now_iso()
+    call_id = f"call_voice_truth_{tmp_path.name}"
+    session_id = f"avs_voice_truth_{tmp_path.name}"
+    ai_session_id = f"ais_voice_truth_{tmp_path.name}"
+    interaction_id = f"int_voice_truth_{tmp_path.name}"
+
+    session = get_session()
+    try:
+        session.add(
+            AsteriskCallLinkRow(
+                call_id=call_id,
+                linked_id=f"linked_voice_truth_{tmp_path.name}",
+                queue_code="voice_lab_ai",
+                queue_id="que_voice_lab",
+                interaction_id=interaction_id,
+                caller_number="+77010002233",
+                caller_name="Truth Caller",
+                status="active",
+                telephony_status="connected",
+                claimed_by_user=None,
+                claimed_at=None,
+                operator_extension=None,
+                channel_name="PJSIP/1001-000111",
+                started_at=now,
+                connected_at=now,
+                ended_at=None,
+                updated_at=now,
+                voice_session_id=session_id,
+                ai_session_id=ai_session_id,
+                ai_state="active",
+                ai_handoff_reason=None,
+                ai_last_model_at=now,
+                voice_start_language="ru",
+            )
+        )
+        session.add(
+            AISessionRow(
+                session_id=ai_session_id,
+                channel="voice",
+                call_id=call_id,
+                thread_id=None,
+                interaction_id=interaction_id,
+                customer_id="cus_voice_truth",
+                agent_profile="voice_support",
+                language="ru",
+                status="active",
+                summary_text="",
+                last_user_message_id=None,
+                last_ai_message_id=None,
+                handoff_reason=None,
+                created_at=now,
+                updated_at=now,
+                closed_at=None,
+            )
+        )
+        session.add(
+            VoiceAISessionRow(
+                session_id=session_id,
+                call_id=call_id,
+                linked_id=f"linked_voice_truth_{tmp_path.name}",
+                interaction_id=interaction_id,
+                customer_id="cus_voice_truth",
+                queue_id="que_voice_lab",
+                ai_session_id=ai_session_id,
+                agent_profile="voice_support",
+                language="ru",
+                asr_provider="elevenlabs",
+                tts_provider="elevenlabs",
+                status="active",
+                handoff_reason=None,
+                handoff_target_queue_id="que_voice_lab",
+                disclosure_played_at=now,
+                last_user_utterance_at=now,
+                last_ai_reply_at=now,
+                started_at=now,
+                updated_at=now,
+                ended_at=None,
+                voice_start_language="ru",
+            )
+        )
+        session.add_all(
+            [
+                VoiceTranscriptSegmentRow(
+                    segment_id=f"{session_id}_seg_1",
+                    session_id=session_id,
+                    call_id=call_id,
+                    interaction_id=interaction_id,
+                    sequence_no=1,
+                    speaker="caller",
+                    source_type="voice_asr",
+                    text="Мне надо узнать график работы.",
+                    confidence=0.96,
+                    is_final=True,
+                    barge_in_interrupted=False,
+                    payload_json='{"intent_bearing": true, "provider": "elevenlabs"}',
+                    created_at=now,
+                ),
+                VoiceTranscriptSegmentRow(
+                    segment_id=f"{session_id}_seg_2",
+                    session_id=session_id,
+                    call_id=call_id,
+                    interaction_id=interaction_id,
+                    sequence_no=2,
+                    speaker="assistant",
+                    source_type="voice_policy",
+                    text="Подскажите, пожалуйста, город или филиал.",
+                    confidence=None,
+                    is_final=True,
+                    barge_in_interrupted=False,
+                    payload_json='{"delivery_state": "delivered", "provider": "elevenlabs"}',
+                    created_at=now,
+                ),
+                VoiceTranscriptSegmentRow(
+                    segment_id=f"{session_id}_seg_3",
+                    session_id=session_id,
+                    call_id=call_id,
+                    interaction_id=interaction_id,
+                    sequence_no=3,
+                    speaker="caller",
+                    source_type="voice_asr",
+                    text="Алло, приветствую.",
+                    confidence=0.8,
+                    is_final=True,
+                    barge_in_interrupted=False,
+                    payload_json='{"intent_bearing": false, "provider": "elevenlabs"}',
+                    created_at=now,
+                ),
+                VoiceTranscriptSegmentRow(
+                    segment_id=f"{session_id}_seg_4",
+                    session_id=session_id,
+                    call_id=call_id,
+                    interaction_id=interaction_id,
+                    sequence_no=4,
+                    speaker="assistant",
+                    source_type="voice_policy",
+                    text="Это прерванный ответ.",
+                    confidence=None,
+                    is_final=True,
+                    barge_in_interrupted=True,
+                    payload_json='{"delivery_state": "interrupted", "provider": "elevenlabs"}',
+                    created_at=now,
+                ),
+            ]
+        )
+        session.commit()
+    finally:
+        session.close()
+
+    summary = bridge_module._voice_ai_summary_for_call(call_id)
+
+    assert summary is not None
+    assert summary.customer_request_text == "Мне надо узнать график работы."
+    assert summary.ai_outcome_text == "Подскажите, пожалуйста, город или филиал."
+    assert [(segment.speaker, segment.text) for segment in summary.transcript_segments] == [
+        ("caller", "Мне надо узнать график работы."),
+        ("assistant", "Подскажите, пожалуйста, город или филиал."),
+        ("caller", "Алло, приветствую."),
+    ]
 def test_call_started_re_raises_original_create_error_when_all_fallbacks_fail(monkeypatch, tmp_path):
    monkeypatch.setenv("ASTERISK_QUEUE_MAP_JSON", '{"lab":"que_lab"}')