move vad config to json

engine/config.py

@@ -28,6 +28,39 @@ class SessionConfig:
     inactivity_timeout_sec: int = 60
 
 
+@dataclass(frozen=True)
+class VADConfig:
+    """Voice Activity Detection thresholds for the Silero analyzer.
+
+    These map directly to ``pipecat.audio.vad.vad_analyzer.VADParams``.
+    Defaults are tuned a touch more conservative than upstream pipecat so
+    short pauses in continuous speech don't end the user turn prematurely.
+    """
+
+    confidence: float = 0.7
+    start_secs: float = 0.2
+    stop_secs: float = 0.6
+    min_volume: float = 0.6
+
+
+@dataclass(frozen=True)
+class TurnConfig:
+    """User-turn segmentation policy.
+
+    ``user_speech_timeout_sec`` is the grace window (in seconds) after VAD
+    has confirmed silence during which the user is allowed to resume
+    speaking before the aggregator finalizes the turn. Used by
+    ``SpeechTimeoutUserTurnStopStrategy``. Higher = more tolerant of
+    natural mid-sentence pauses; lower = snappier turn-taking.
+
+    The combined "user pause before turn ends" budget is roughly
+    ``vad.stop_secs + user_speech_timeout_sec``.
+    """
+
+    vad: VADConfig = field(default_factory=VADConfig)
+    user_speech_timeout_sec: float = 1.0
+
+
 @dataclass(frozen=True)
 class AgentConfig:
     system_prompt: str = "You are a helpful, friendly voice assistant."
@@ -90,6 +123,7 @@ class EngineConfig:
     server: ServerConfig = field(default_factory=ServerConfig)
     audio: AudioConfig = field(default_factory=AudioConfig)
     session: SessionConfig = field(default_factory=SessionConfig)
+    turn: TurnConfig = field(default_factory=TurnConfig)
     agent: AgentConfig = field(default_factory=AgentConfig)
     services: ServicesConfig = field(default_factory=ServicesConfig)
 
@@ -116,10 +150,19 @@ def config_from_dict(data: dict) -> EngineConfig:
     if stt.get("language") == "":
         stt["language"] = None
 
+    turn = _dict(data.get("turn"))
+    vad = _dict(turn.get("vad"))
+
     return EngineConfig(
         server=ServerConfig(**_dict(data.get("server"))),
         audio=AudioConfig(**_dict(data.get("audio"))),
         session=SessionConfig(**_dict(data.get("session"))),
+        turn=TurnConfig(
+            vad=VADConfig(**vad),
+            user_speech_timeout_sec=float(
+                turn.get("user_speech_timeout_sec", TurnConfig().user_speech_timeout_sec)
+            ),
+        ),
         agent=AgentConfig(**agent),
         services=ServicesConfig(
             llm=LLMConfig(**_dict(services.get("llm"))),