세션 메모리와 비언어 이벤트 저장

apps/api/app/routes/voice.py

@@ -48,6 +48,56 @@ WS_CLOSE_UNAUTHORIZED = 1008
 
 # Per-utterance audio cap to avoid unbounded memory growth.
 _MAX_AUDIO_BYTES = 10 * 1024 * 1024
+_PROVIDER_EVENT_MAX_ITEMS = 12
+_PROVIDER_EVENT_MAX_STRING = 80
+_PROVIDER_EVENT_ALLOWED_KEYS = {
+    "category",
+    "event_type",
+    "type",
+    "kind",
+    "label",
+    "source",
+    "provider",
+    "start_ms",
+    "end_ms",
+    "duration_ms",
+    "confidence",
+    "score",
+    "is_final",
+}
+_PROVIDER_EVENT_TAXONOMY = {
+    "barge_in": ("barge_in", "turn_taking"),
+    "interrupt": ("barge_in", "turn_taking"),
+    "interruption": ("barge_in", "turn_taking"),
+    "overlap": ("barge_in", "turn_taking"),
+    "sigh": ("sigh", "paralinguistic"),
+    "sighing": ("sigh", "paralinguistic"),
+    "sob": ("cry", "paralinguistic"),
+    "cry": ("cry", "paralinguistic"),
+    "crying": ("cry", "paralinguistic"),
+    "weep": ("cry", "paralinguistic"),
+    "laugh": ("laugh", "paralinguistic"),
+    "laughter": ("laugh", "paralinguistic"),
+    "breath": ("breath", "paralinguistic"),
+    "breathing": ("breath", "paralinguistic"),
+    "voice_activity": ("voice_activity", "speech_activity"),
+    "vad": ("voice_activity", "speech_activity"),
+    "speech_start": ("speech_start", "speech_activity"),
+    "speech_end": ("speech_end", "speech_activity"),
+    "speech_final": ("speech_final", "speech_activity"),
+    "silence": ("silence", "timing"),
+    "pause": ("silence", "timing"),
+    "long_pause": ("silence", "timing"),
+    "speech_rate": ("speech_rate", "prosody"),
+    "fast_speech": ("speech_rate", "prosody"),
+    "slow_speech": ("speech_rate", "prosody"),
+    "pitch": ("pitch", "prosody"),
+    "intonation": ("intonation", "prosody"),
+    "prosody": ("prosody", "prosody"),
+    "noise": ("background_noise", "audio_quality"),
+    "background_noise": ("background_noise", "audio_quality"),
+}
+_PROVIDER_EVENT_TYPE_FIELDS = ("event_type", "type", "kind", "label")
 
 
 @router.get("/health")
@@ -187,6 +237,7 @@ async def voice_ws(websocket: WebSocket) -> None:
                     audio_ended_at=audio_ended_at,
                     silence_ms=silence_ms,
                     barge_in=_safe_bool(ctrl.get("barge_in")),
+                    provider_events=_safe_provider_events(ctrl.get("provider_events")),
                 )
                 last_audio_end_at = audio_ended_at
                 audio_started_at = None
@@ -232,6 +283,7 @@ async def _handle_utterance(
     audio_ended_at: float | None = None,
     silence_ms: int | None = None,
     barge_in: bool | None = None,
+    provider_events: list[dict[str, object]] | None = None,
 ) -> None:
     """Transcribe one utterance, generate the client reply, then synthesize TTS."""
     if not audio:
@@ -259,6 +311,7 @@ async def _handle_utterance(
     audio_ref = _voice_audio_ref(audio, fmt)
     duration_s = stt.duration or _elapsed_seconds(audio_started_at, audio_ended_at)
     speech_rate = _estimate_speech_rate(learner_text, duration_s)
+    provider_events = _merge_provider_events(provider_events, getattr(stt, "provider_events", []))
     await _safe_send_json(
         websocket,
         {"type": "transcript", "text": learner_text, "final": True, "speaker": "counselor"},
@@ -277,6 +330,7 @@ async def _handle_utterance(
         silence_ms=silence_ms,
         speech_rate=speech_rate,
         barge_in=barge_in,
+        provider_events=provider_events,
     )
 
 
@@ -291,6 +345,7 @@ async def _run_turn_and_speak(
     silence_ms: int | None = None,
     speech_rate: float | None = None,
     barge_in: bool | None = None,
+    provider_events: list[dict[str, object]] | None = None,
 ) -> None:
     """Run one counseling turn and stream synthesized client speech."""
     sess, err = await _load_voice_session(session_id, principal)
@@ -351,6 +406,7 @@ async def _run_turn_and_speak(
             silence_ms=silence_ms,
             speech_rate=speech_rate,
             barge_in=barge_in,
+            provider_events=provider_events or [],
             evaluation=result.evaluation,
         ),
     )
@@ -646,6 +702,71 @@ def _safe_bool(value: object) -> bool | None:
     return bool(value)
 
 
+def _provider_event_slug(value: object) -> str:
+    text = str(value or "").strip().lower()
+    if not text:
+        return ""
+    chars = []
+    previous_underscore = False
+    for char in text:
+        if char.isalnum():
+            chars.append(char)
+            previous_underscore = False
+        elif not previous_underscore:
+            chars.append("_")
+            previous_underscore = True
+    return "".join(chars).strip("_")[:_PROVIDER_EVENT_MAX_STRING]
+
+
+def _provider_event_taxonomy(event: dict[str, object]) -> tuple[str, str]:
+    for field in _PROVIDER_EVENT_TYPE_FIELDS:
+        slug = _provider_event_slug(event.get(field))
+        if not slug:
+            continue
+        canonical = _PROVIDER_EVENT_TAXONOMY.get(slug)
+        if canonical:
+            return canonical
+        return slug, "unknown"
+    return "", ""
+
+
+def _safe_provider_events(value: object) -> list[dict[str, object]]:
+    if not isinstance(value, list):
+        return []
+    events: list[dict[str, object]] = []
+    for item in value[:_PROVIDER_EVENT_MAX_ITEMS]:
+        if not isinstance(item, dict):
+            continue
+        safe: dict[str, object] = {}
+        for key in _PROVIDER_EVENT_ALLOWED_KEYS:
+            raw = item.get(key)
+            if isinstance(raw, bool):
+                safe[key] = raw
+            elif isinstance(raw, (int, float)):
+                safe[key] = raw
+            elif isinstance(raw, str):
+                text = raw.strip()
+                if text:
+                    safe[key] = text[:_PROVIDER_EVENT_MAX_STRING]
+        if safe:
+            event_type, category = _provider_event_taxonomy(safe)
+            if event_type:
+                safe["event_type"] = event_type
+            if category:
+                safe["category"] = category
+            events.append(safe)
+    return events
+
+
+def _merge_provider_events(*values: object) -> list[dict[str, object]]:
+    merged: list[dict[str, object]] = []
+    for value in values:
+        merged.extend(_safe_provider_events(value))
+        if len(merged) >= _PROVIDER_EVENT_MAX_ITEMS:
+            return merged[:_PROVIDER_EVENT_MAX_ITEMS]
+    return merged
+
+
 async def _safe_send_json(websocket: WebSocket, payload: dict) -> None:
     if websocket.client_state != WebSocketState.CONNECTED:
         return