dv-pipecat-ai 0.0.85.dev844__py3-none-any.whl → 0.0.85.dev848__py3-none-any.whl

{dv_pipecat_ai-0.0.85.dev844.dist-info → dv_pipecat_ai-0.0.85.dev848.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: dv-pipecat-ai
-Version: 0.0.85.dev844
+Version: 0.0.85.dev848
 Summary: An open source framework for voice (and multimodal) assistants
 License-Expression: BSD-2-Clause
 Project-URL: Source, https://github.com/pipecat-ai/pipecat

{dv_pipecat_ai-0.0.85.dev844.dist-info → dv_pipecat_ai-0.0.85.dev848.dist-info}/RECORD

@@ -1,4 +1,4 @@
-dv_pipecat_ai-0.0.85.dev844.dist-info/licenses/LICENSE,sha256=DWY2QGf2eMCFhuu2ChairtT6CB7BEFffNVhXWc4Od08,1301
+dv_pipecat_ai-0.0.85.dev848.dist-info/licenses/LICENSE,sha256=DWY2QGf2eMCFhuu2ChairtT6CB7BEFffNVhXWc4Od08,1301
 pipecat/__init__.py,sha256=j0Xm6adxHhd7D06dIyyPV_GlBYLlBnTAERVvD_jAARQ,861
 pipecat/py.typed,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 pipecat/adapters/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
@@ -79,7 +79,7 @@ pipecat/extensions/voicemail/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NM
 pipecat/extensions/voicemail/voicemail_detector.py,sha256=JxmU2752iWP_1_GmzZReNESUTFAeyEa4XBPL20_C208,30004
 pipecat/frames/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 pipecat/frames/frames.proto,sha256=JXZm3VXLR8zMOUcOuhVoe2mhM3MQIQGMJXLopdJO_5Y,839
-pipecat/frames/frames.py,sha256=vuYtmyK1QSU2AWx2c_pFQhcmpXqSTnfqAXF6DXKzTG8,49605
+pipecat/frames/frames.py,sha256=248d54lNOyO04dq9ni51yUTWUItmGw8b9QKarrDGNeo,50354
 pipecat/frames/protobufs/frames_pb2.py,sha256=VHgGV_W7qQ4sfQK6RHb5_DggLm3PiSYMr6aBZ8_p1cQ,2590
 pipecat/metrics/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 pipecat/metrics/metrics.py,sha256=bdZNciEtLTtA-xgoKDz2RJAy6fKrXkTwz3pryVHzc2M,2713
@@ -217,8 +217,8 @@ pipecat/services/deepgram/flux/stt.py,sha256=yCZodrHAOShgYy_GbdviX8iAuh36dBgDL41
 pipecat/services/deepseek/__init__.py,sha256=bU5z_oNGzgrF_YpsD9pYIMtEibeZFaUobbRjJ9WcYyE,259
 pipecat/services/deepseek/llm.py,sha256=5KjpU2blmhUTM3LcRE1ymdsk6OmoFkIzeQgyNOGwQh8,3112
 pipecat/services/elevenlabs/__init__.py,sha256=cMx5v0HEMh4WetMm5byR9tIjG6_wNVs9UxqWyB3tjlM,313
-pipecat/services/elevenlabs/stt.py,sha256=F3xD82eOIy5OyyE-5StdoFFvKjIXlos2yyP0cyNQj6Y,12214
-pipecat/services/elevenlabs/tts.py,sha256=Okctydqoz2HG9B69l-bDunNHcqoul-kUxhEZjYt539U,45188
+pipecat/services/elevenlabs/stt.py,sha256=_RhBKpUYEGKMpcO7y4RLxmEOMK11LZFdZqDFIA-DZXk,27303
+pipecat/services/elevenlabs/tts.py,sha256=skUndgUatx2F5rjg2tBZLutB8k9B9Cjy-cUeglCDdwc,45314
 pipecat/services/fal/__init__.py,sha256=z_kfZETvUcKy68Lyvni4B-RtdkOvz3J3eh6sFDVKq6M,278
 pipecat/services/fal/image.py,sha256=vArKLKrIGoZfw_xeZY_E7zbUzfzVsScj-R7mOmVqjRQ,4585
 pipecat/services/fal/stt.py,sha256=-5tw7N8srBJTS0Q65SN4csmLkIB6cLHR9pXKimxg55o,9678
@@ -280,7 +280,7 @@ pipecat/services/nim/llm.py,sha256=o4WPGI6kOmSiMV7WwOZ0cNEAoq9hW4Aqs2R8X7c9i94,4
 pipecat/services/ollama/__init__.py,sha256=aw-25zYsR8LR74OFFlMKMTnJjaKwOzdPWVsClueNRkI,255
 pipecat/services/ollama/llm.py,sha256=rfpG92LRHGJlpENKhF6ld8CLVS9DxlKW-WRVNldOIGs,1605
 pipecat/services/openai/__init__.py,sha256=V0ZVa8PzEm3hmcStYICbAsYwfgk4ytZ6kiQoq9UZPmI,354
-pipecat/services/openai/base_llm.py,sha256=J4Ltg1KOXciiUIMBFLn0SmDTZereEE-1LKrPfBsLzFw,19127
+pipecat/services/openai/base_llm.py,sha256=jOiWacimREywCMZZwAwH8RAHCbwnnXvbqAjWQUYA0yM,20727
 pipecat/services/openai/image.py,sha256=3e3h-dVQ6DQuQE7fp8akXwRMd-oYOdGuZg7RCOjHu9A,2994
 pipecat/services/openai/llm.py,sha256=_aKtz1VebSFUUenT3tH6mBW9pSCm65_u45cDu_dkTzs,7396
 pipecat/services/openai/stt.py,sha256=Idf0k73kxFyDgNRBt62MFpoKKNsBV9bwvJteJ6MGWzQ,2419
@@ -329,7 +329,7 @@ pipecat/services/sarvam/tts.py,sha256=lrwfdC53kZ7f2QPgNRxzryISNkrJCvNtlZ-19-iXg9
 pipecat/services/simli/__init__.py,sha256=cbDcqOaGsEgKbGYKpJ1Vv7LN4ZjOWA04sE84WW5vgQI,257
 pipecat/services/simli/video.py,sha256=Zu2XLvl2Y6VHaWzT9wEdzW9d0EYoZyzYLxjQFyV8vho,8320
 pipecat/services/soniox/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-pipecat/services/soniox/stt.py,sha256=Ndml6QvPQ1WZBvdGT3LSg-LLWwrZ8KlqW8wBBFsQrrM,16509
+pipecat/services/soniox/stt.py,sha256=zRp5qWU051hEAikt0vB0rbHrkQkH5sT-IOe-o5vCurQ,16650
 pipecat/services/speechmatics/__init__.py,sha256=Jgq1fqrZVkpWC21D79L1cn5Ub8PnYgnnCaqC5pOlbIc,89
 pipecat/services/speechmatics/stt.py,sha256=GLGJzlMSeZ1WzTOMjhKXDl5JYkqGhnFTbP3o0ez0hSw,44696
 pipecat/services/tavus/__init__.py,sha256=SNyyi2Xq6tXIihDG2Bwvmg6Srbd-uWd1RwG-NKWcPuI,257
@@ -416,7 +416,7 @@ pipecat/utils/tracing/service_decorators.py,sha256=fwzxFpi8DJl6BJbK74G0UEB4ccMJg
 pipecat/utils/tracing/setup.py,sha256=7TEgPNpq6M8lww8OQvf0P9FzYc5A30xICGklVA-fua0,2892
 pipecat/utils/tracing/turn_context_provider.py,sha256=ikon3plFOx0XbMrH6DdeHttNpb-U0gzMZIm3bWLc9eI,2485
 pipecat/utils/tracing/turn_trace_observer.py,sha256=dma16SBJpYSOE58YDWy89QzHyQFc_9gQZszKeWixuwc,9725
-dv_pipecat_ai-0.0.85.dev844.dist-info/METADATA,sha256=JgW9PLS_gplsOlHfyohgocRxrsiivvsAEySMY214f4U,32955
-dv_pipecat_ai-0.0.85.dev844.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-dv_pipecat_ai-0.0.85.dev844.dist-info/top_level.txt,sha256=kQzG20CxGf-nSsHmtXHx3hY2-8zHA3jYg8jk0TajqXc,8
-dv_pipecat_ai-0.0.85.dev844.dist-info/RECORD,,
+dv_pipecat_ai-0.0.85.dev848.dist-info/METADATA,sha256=T2IPoO2Nmt09lIxM0rKmJRa5ZIBQ-9fcbswOy90lkJg,32955
+dv_pipecat_ai-0.0.85.dev848.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+dv_pipecat_ai-0.0.85.dev848.dist-info/top_level.txt,sha256=kQzG20CxGf-nSsHmtXHx3hY2-8zHA3jYg8jk0TajqXc,8
+dv_pipecat_ai-0.0.85.dev848.dist-info/RECORD,,

pipecat/frames/frames.py

@@ -586,6 +586,27 @@ class LLMRunFrame(DataFrame):
     pass
 
 
+@dataclass
+class WarmupLLMFrame(DataFrame):
+    """Frame to trigger prompt caching/warmup in supported LLM providers.
+
+    This frame instructs the LLM service to cache the provided messages
+    without generating a visible response. Primarily used for warming up provider
+    caches (e.g., Claude's prompt caching, OpenAI's prompt caching) to improve
+    latency for subsequent requests.
+
+    The LLM service should:
+    1. Send the messages to the provider to trigger caching
+    2. Generate a minimal response (e.g., single word)
+    3. Discard the response without emitting LLM output frames
+
+    Parameters:
+        messages: List of messages to send for cache warming (should match conversation structure).
+    """
+
+    messages: List[dict]
+
+
 @dataclass
 class LLMMessagesAppendFrame(DataFrame):
     """Frame containing LLM messages to append to current context.

pipecat/services/elevenlabs/stt.py

@@ -4,26 +4,43 @@
 # SPDX-License-Identifier: BSD 2-Clause License
 #
 
-"""ElevenLabs speech-to-text service implementation.
-
-This module provides integration with ElevenLabs' Speech-to-Text API for transcription
-using segmented audio processing. The service uploads audio files and receives
-transcription results directly.
-"""
+"""ElevenLabs speech-to-text service implementations."""
 
+import asyncio
+import base64
 import io
-from typing import AsyncGenerator, Optional
+import json
+import urllib.parse
+from typing import Any, AsyncGenerator, Dict, Literal, Optional
 
 import aiohttp
 from loguru import logger
 from pydantic import BaseModel
 
-from pipecat.frames.frames import ErrorFrame, Frame, TranscriptionFrame
-from pipecat.services.stt_service import SegmentedSTTService
+from pipecat.frames.frames import (
+    CancelFrame,
+    EndFrame,
+    ErrorFrame,
+    Frame,
+    InterimTranscriptionFrame,
+    StartFrame,
+    TranscriptionFrame,
+    UserStartedSpeakingFrame,
+    UserStoppedSpeakingFrame,
+)
+from pipecat.processors.frame_processor import FrameDirection
+from pipecat.services.stt_service import SegmentedSTTService, WebsocketSTTService
 from pipecat.transcriptions.language import Language
 from pipecat.utils.time import time_now_iso8601
 from pipecat.utils.tracing.service_decorators import traced_stt
 
+try:
+    from websockets.asyncio.client import connect as websocket_connect
+    from websockets.protocol import State
+except ModuleNotFoundError:
+    websocket_connect = None  # type: ignore[assignment]
+    State = None  # type: ignore[assignment]
+
 
 def language_to_elevenlabs_language(language: Language) -> Optional[str]:
     """Convert a Language enum to ElevenLabs language code.
@@ -150,6 +167,19 @@ def language_to_elevenlabs_language(language: Language) -> Optional[str]:
     return result
 
 
+def elevenlabs_language_code_to_language(language_code: Optional[str]) -> Optional[Language]:
+    """Convert an ElevenLabs language code back to a Language enum value."""
+    if not language_code:
+        return None
+
+    normalized = language_code.lower()
+    for language in Language:
+        code = language_to_elevenlabs_language(language)
+        if code and code.lower() == normalized:
+            return language
+    return None
+
+
 class ElevenLabsSTTService(SegmentedSTTService):
     """Speech-to-text service using ElevenLabs' file-based API.
 
@@ -337,3 +367,376 @@ class ElevenLabsSTTService(SegmentedSTTService):
         except Exception as e:
             self.logger.error(f"ElevenLabs STT error: {e}")
             yield ErrorFrame(f"ElevenLabs STT error: {str(e)}")
+
+
+class ElevenLabsRealtimeSTTService(WebsocketSTTService):
+    """Realtime speech-to-text service using ElevenLabs Scribe v2 WebSocket API."""
+
+    class InputParams(BaseModel):
+        """Realtime connection parameters derived from ElevenLabs documentation."""
+
+        language: Optional[Language] = None
+        commit_strategy: Literal["manual", "vad"] = "manual"
+        vad_silence_threshold_secs: Optional[float] = None
+        vad_threshold: Optional[float] = None
+        min_speech_duration_ms: Optional[int] = None
+        min_silence_duration_ms: Optional[int] = None
+
+    def __init__(
+        self,
+        *,
+        api_key: str,
+        sample_rate: Optional[int] = None,
+        model: str = "scribe_v2_realtime",
+        url: str = "wss://api.elevenlabs.io/v1/speech-to-text/realtime",
+        params: Optional["ElevenLabsRealtimeSTTService.InputParams"] = None,
+        reconnect_on_error: bool = True,
+        **kwargs,
+    ):
+        """Initialize the realtime STT service.
+
+        Args:
+            api_key: ElevenLabs API key for authentication.
+            sample_rate: Optional input sample rate. Defaults to pipeline sample rate.
+            model: Scribe realtime model identifier.
+            url: WebSocket endpoint for realtime transcription.
+            params: Optional realtime configuration options.
+            reconnect_on_error: Whether to auto-reconnect on transient failures.
+            **kwargs: Additional arguments forwarded to WebsocketSTTService.
+        """
+        if websocket_connect is None or State is None:
+            logger.error(
+                "In order to use ElevenLabsRealtimeSTTService, you need to "
+                "`pip install pipecat-ai[elevenlabs]` (websockets extra)."
+            )
+            raise ModuleNotFoundError("Missing optional dependency: websockets")
+
+        super().__init__(sample_rate=sample_rate, reconnect_on_error=reconnect_on_error, **kwargs)
+
+        self._api_key = api_key
+        self._url = url
+        self.set_model_name(model)
+        self._model = model
+        self._params = params or ElevenLabsRealtimeSTTService.InputParams()
+        self._language_override = self._params.language
+        self._encoding = None
+        self._receive_task: Optional[asyncio.Task] = None
+        self._pending_final_message: Optional[Dict[str, Any]] = None
+        self._pending_final_task: Optional[asyncio.Task] = None
+        self._timestamp_merge_delay_s = 0.25
+        self._ttfb_started = False
+
+    @property
+    def commit_strategy(self) -> str:
+        """Return the configured commit strategy (manual or vad)."""
+        return (self._params.commit_strategy or "manual").lower()
+
+    def can_generate_metrics(self) -> bool:
+        """Realtime ElevenLabs service supports latency metrics."""
+        return True
+
+    async def start(self, frame: StartFrame):
+        """Start the realtime STT service and establish WebSocket connection."""
+        await super().start(frame)
+        self._encoding = self._determine_encoding(self.sample_rate)
+        await self._connect()
+
+    async def stop(self, frame: EndFrame):
+        """Stop the realtime STT service and close WebSocket connection."""
+        await super().stop(frame)
+        await self._disconnect()
+
+    async def cancel(self, frame: CancelFrame):
+        """Cancel the realtime STT service and close WebSocket connection."""
+        await super().cancel(frame)
+        await self._disconnect()
+
+    async def set_language(self, language: Language):
+        """Update preferred transcription language (requires reconnect)."""
+        self._language_override = language
+        self._params.language = language
+        if self._websocket:
+            await self._disconnect()
+            await self._connect()
+
+    async def set_model(self, model: str):
+        """Set the STT model and reconnect the WebSocket."""
+        await super().set_model(model)
+        self._model = model
+        if self._websocket:
+            await self._disconnect()
+            await self._connect()
+
+    async def process_frame(self, frame: Frame, direction: FrameDirection):
+        """Process frames and handle VAD events for commit strategy."""
+        await super().process_frame(frame, direction)
+
+        if isinstance(frame, UserStartedSpeakingFrame):
+            if frame.emulated:
+                return
+            self._ttfb_started = False
+            await self.start_processing_metrics()
+        elif isinstance(frame, UserStoppedSpeakingFrame):
+            if frame.emulated:
+                return
+            if self.commit_strategy == "manual":
+                await self._send_commit()
+
+    async def run_stt(self, audio: bytes) -> AsyncGenerator[Frame, None]:
+        """Stream audio chunks over the ElevenLabs realtime WebSocket."""
+        if not audio:
+            yield None
+            return
+
+        await self._ensure_connection()
+        await self._send_audio_chunk(audio)
+        yield None
+
+    async def _ensure_connection(self):
+        if not self._websocket or self._websocket.state is State.CLOSED:
+            await self._connect()
+
+    async def _connect(self):
+        await self._connect_websocket()
+        if self._websocket and not self._receive_task:
+            self._receive_task = asyncio.create_task(self._receive_task_handler(self._report_error))
+
+    async def _disconnect(self):
+        if self._receive_task:
+            await self.cancel_task(self._receive_task)
+            self._receive_task = None
+
+        await self._clear_pending_final()
+        await self._disconnect_websocket()
+
+    async def _connect_websocket(self):
+        try:
+            if self._websocket and self._websocket.state is State.OPEN:
+                return
+
+            ws_url = self._build_websocket_url()
+            headers = {"xi-api-key": self._api_key}
+            self.logger.debug(f"Connecting to ElevenLabs realtime STT at {ws_url}")
+            self._websocket = await websocket_connect(ws_url, additional_headers=headers)
+            await self._call_event_handler("on_connected")
+        except Exception as e:
+            self.logger.error(f"{self} unable to connect to ElevenLabs realtime STT: {e}")
+            self._websocket = None
+            await self._call_event_handler("on_connection_error", f"{e}")
+
+    async def _disconnect_websocket(self):
+        try:
+            await self.stop_all_metrics()
+            if self._websocket and self._websocket.state is State.OPEN:
+                self.logger.debug("Disconnecting from ElevenLabs realtime STT")
+                await self._websocket.close()
+        except Exception as e:
+            self.logger.error(f"{self} error closing ElevenLabs realtime websocket: {e}")
+        finally:
+            self._websocket = None
+            await self._call_event_handler("on_disconnected")
+
+    async def _receive_messages(self):
+        async for message in self._get_websocket():
+            await self._process_event(message)
+
+    def _get_websocket(self):
+        if not self._websocket:
+            raise RuntimeError("ElevenLabs realtime websocket not connected")
+        return self._websocket
+
+    async def _process_event(self, message: Any):
+        try:
+            data = json.loads(message)
+        except json.JSONDecodeError:
+            self.logger.warning(f"ElevenLabs realtime STT sent invalid JSON: {message}")
+            return
+
+        message_type = data.get("message_type")
+
+        if message_type == "session_started":
+            self.logger.debug("ElevenLabs realtime session started")
+            return
+
+        if message_type == "partial_transcript":
+            await self._emit_partial_transcript(data)
+        elif message_type == "committed_transcript":
+            await self._handle_committed_transcript(data)
+        elif message_type == "committed_transcript_with_timestamps":
+            await self._handle_committed_transcript_with_timestamps(data)
+        elif message_type in {
+            "auth_error",
+            "quota_exceeded",
+            "transcriber_error",
+            "input_error",
+            "error",
+        }:
+            fatal = message_type in {"auth_error", "quota_exceeded", "error"}
+            description = data.get("error", data)
+            await self.push_error(
+                ErrorFrame(f"ElevenLabs realtime error: {description}", fatal=fatal)
+            )
+        else:
+            self.logger.debug(f"Unhandled ElevenLabs realtime message: {data}")
+
+    async def _emit_partial_transcript(self, data: Dict[str, Any]):
+        text = (data.get("text") or data.get("transcript") or "").strip()
+        if not text:
+            return
+
+        language = (
+            elevenlabs_language_code_to_language(data.get("language_code"))
+            or self._language_override
+        )
+        await self.stop_ttfb_metrics()
+
+        await self.push_frame(
+            InterimTranscriptionFrame(
+                text,
+                self._user_id,
+                time_now_iso8601(),
+                language,
+                result=data,
+            )
+        )
+
+    async def _handle_committed_transcript(self, data: Dict[str, Any]):
+        if self._pending_final_message:
+            await self._emit_transcription(self._pending_final_message)
+            self._pending_final_message = None
+
+        self._pending_final_message = data
+        await self._schedule_pending_final_emit()
+
+    async def _handle_committed_transcript_with_timestamps(self, data: Dict[str, Any]):
+        if self._pending_final_message:
+            merged = {**self._pending_final_message, **data}
+            await self._emit_transcription(merged)
+            await self._clear_pending_final()
+        else:
+            await self._emit_transcription(data)
+
+    async def _schedule_pending_final_emit(self):
+        await self._clear_pending_final(timer_only=True)
+        self._pending_final_task = asyncio.create_task(self._emit_pending_after_delay())
+
+    async def _emit_pending_after_delay(self):
+        try:
+            await asyncio.sleep(self._timestamp_merge_delay_s)
+            if self._pending_final_message:
+                await self._emit_transcription(self._pending_final_message)
+                self._pending_final_message = None
+        except asyncio.CancelledError:
+            pass
+        finally:
+            self._pending_final_task = None
+
+    async def _clear_pending_final(self, timer_only: bool = False):
+        if self._pending_final_task:
+            await self.cancel_task(self._pending_final_task)
+            self._pending_final_task = None
+
+        if not timer_only:
+            self._pending_final_message = None
+
+    async def _emit_transcription(self, data: Dict[str, Any]):
+        text = (data.get("text") or data.get("transcript") or "").strip()
+        if not text:
+            return
+
+        language = (
+            elevenlabs_language_code_to_language(data.get("language_code"))
+            or self._language_override
+        )
+        await self.stop_ttfb_metrics()
+
+        frame = TranscriptionFrame(
+            text,
+            self._user_id,
+            time_now_iso8601(),
+            language,
+            result=data,
+        )
+
+        await self.push_frame(frame)
+        await self._handle_transcription(text, True, language)
+        await self.stop_processing_metrics()
+
+    async def _send_audio_chunk(self, audio: bytes):
+        if not audio or not self._websocket:
+            return
+
+        if not self._ttfb_started:
+            await self.start_ttfb_metrics()
+            self._ttfb_started = True
+
+        payload = {
+            "message_type": "input_audio_chunk",
+            "audio_base_64": base64.b64encode(audio).decode("ascii"),
+            "commit": False,
+            "sample_rate": self.sample_rate,
+        }
+        await self._websocket.send(json.dumps(payload))
+
+    async def _send_commit(self):
+        if not self._websocket:
+            return
+        payload = {
+            "message_type": "input_audio_chunk",
+            "audio_base_64": "",
+            "commit": True,
+            "sample_rate": self.sample_rate,
+        }
+        await self._websocket.send(json.dumps(payload))
+
+    def _build_websocket_url(self) -> str:
+        if not self.sample_rate:
+            raise ValueError(
+                "ElevenLabs realtime STT requires a valid sample rate (start() must run first)."
+            )
+
+        params = {
+            "model_id": self._model,
+            "encoding": self._encoding or "pcm_16000",
+            "sample_rate": str(self.sample_rate),
+            "commit_strategy": self.commit_strategy,
+        }
+
+        language_code = (
+            language_to_elevenlabs_language(self._language_override)
+            if self._language_override
+            else None
+        )
+        if language_code:
+            params["language_code"] = language_code
+
+        if self._params.vad_silence_threshold_secs is not None:
+            params["vad_silence_threshold_secs"] = str(self._params.vad_silence_threshold_secs)
+        if self._params.vad_threshold is not None:
+            params["vad_threshold"] = str(self._params.vad_threshold)
+        if self._params.min_speech_duration_ms is not None:
+            params["min_speech_duration_ms"] = str(self._params.min_speech_duration_ms)
+        if self._params.min_silence_duration_ms is not None:
+            params["min_silence_duration_ms"] = str(self._params.min_silence_duration_ms)
+
+        return f"{self._url}?{urllib.parse.urlencode(params)}"
+
+    def _determine_encoding(self, sample_rate: int) -> str:
+        if not sample_rate:
+            raise ValueError("ElevenLabs realtime STT requires a valid sample rate.")
+
+        supported_rates = {8000, 16000, 22050, 24000, 44100, 48000}
+        if sample_rate not in supported_rates:
+            raise ValueError(
+                f"ElevenLabs realtime STT supports sample rates {sorted(supported_rates)}. "
+                f"Received {sample_rate} Hz."
+            )
+        return f"pcm_{sample_rate}"
+
+    @traced_stt
+    async def _handle_transcription(
+        self, transcript: str, is_final: bool, language: Optional[Language] = None
+    ):
+        """Handle a transcription result with tracing."""
+        # Metrics are stopped by the caller when needed.
+        return

pipecat/services/elevenlabs/tts.py

@@ -14,7 +14,17 @@ import asyncio
 import base64
 import json
 import uuid
-from typing import Any, AsyncGenerator, Dict, List, Literal, Mapping, Optional, Tuple, Union
+from typing import (
+    Any,
+    AsyncGenerator,
+    Dict,
+    List,
+    Literal,
+    Mapping,
+    Optional,
+    Tuple,
+    Union,
+)
 
 import aiohttp
 from loguru import logger
@@ -157,7 +167,13 @@ def build_elevenlabs_voice_settings(
     Returns:
         Dictionary of voice settings or None if no valid settings are provided.
     """
-    voice_setting_keys = ["stability", "similarity_boost", "style", "use_speaker_boost", "speed"]
+    voice_setting_keys = [
+        "stability",
+        "similarity_boost",
+        "style",
+        "use_speaker_boost",
+        "speed",
+    ]
 
     voice_settings = {}
     for key in voice_setting_keys:
@@ -530,7 +546,9 @@ class ElevenLabsTTSService(AudioContextWordTTSService):
 
             # Set max websocket message size to 16MB for large audio responses
             self._websocket = await websocket_connect(
-                url, max_size=16 * 1024 * 1024, additional_headers={"xi-api-key": self._api_key}
+                url,
+                max_size=16 * 1024 * 1024,
+                additional_headers={"xi-api-key": self._api_key},
             )
 
             await self._call_event_handler("on_connected")
@@ -549,7 +567,7 @@ class ElevenLabsTTSService(AudioContextWordTTSService):
                 if self._context_id:
                     await self._websocket.send(json.dumps({"close_socket": True}))
                 await self._websocket.close()
-                logger.debug("Disconnected from ElevenLabs")
+                self.logger.debug("Disconnected from ElevenLabs")
         except Exception as e:
             self.logger.error(f"{self} error closing websocket: {e}")
         finally:

pipecat/services/openai/base_llm.py

@@ -32,6 +32,7 @@ from pipecat.frames.frames import (
     LLMMessagesFrame,
     LLMTextFrame,
     LLMUpdateSettingsFrame,
+    WarmupLLMFrame,
 )
 from pipecat.metrics.metrics import LLMTokenUsage
 from pipecat.processors.aggregators.llm_context import LLMContext
@@ -438,14 +439,19 @@ class BaseOpenAILLMService(LLMService):
                 completions and manage settings.
         >>>>>>> dv-stage
 
-                Args:
+        Args:
                     frame: The frame to process.
                     direction: The direction of frame processing.
         """
         await super().process_frame(frame, direction)
 
         context = None
-        if isinstance(frame, OpenAILLMContextFrame):
+        if isinstance(frame, WarmupLLMFrame):
+            # Handle warmup frame - prime cache without emitting response
+            # Run in background to avoid blocking the pipeline
+            asyncio.create_task(self._handle_warmup_frame(frame))
+            return  # Don't process further, warmup is silent
+        elif isinstance(frame, OpenAILLMContextFrame):
             # Handle OpenAI-specific context frames
             context = frame.context
         elif isinstance(frame, LLMContextFrame):
@@ -470,3 +476,32 @@ class BaseOpenAILLMService(LLMService):
             finally:
                 await self.stop_processing_metrics()
                 await self.push_frame(LLMFullResponseEndFrame())
+
+    async def _handle_warmup_frame(self, frame: WarmupLLMFrame):
+        """Handle WarmupLLMFrame to prime the LLM cache without emitting responses.
+
+        This method sends a minimal request to the LLM to warm up any provider-side
+        caches (like prompt caching). The response is discarded and no frames are emitted.
+
+        Args:
+            frame: WarmupLLMFrame containing the messages to cache.
+        """
+        try:
+            # Use the provided messages for warmup
+            messages: List[ChatCompletionMessageParam] = frame.messages  # type: ignore
+
+            # Make a non-streaming call to warm the cache
+            # We use a minimal max_tokens to reduce latency and cost
+            await self._client.chat.completions.create(
+                model=self.model_name,  # Use the property, not self._model
+                messages=messages,
+                max_tokens=10,  # Minimal response
+                stream=False,
+            )
+
+            self.logger.info("LLM cache warmed successfully")
+            # Intentionally don't emit any frames - this is a silent warmup
+
+        except Exception as e:
+            self.logger.error(f"Failed to warm LLM cache: {e}")
+            # Don't propagate error - warmup failure shouldn't break the bot

pipecat/services/soniox/stt.py

@@ -241,6 +241,7 @@ class SonioxSTTService(STTService):
             if self._receive_task != asyncio.current_task():
                 await self._receive_task
             self._receive_task = None
+        self.logger.debug("Disconnected from Soniox STT")
 
     async def stop(self, frame: EndFrame):
         """Stop the Soniox STT websocket connection.
@@ -382,7 +383,10 @@ class SonioxSTTService(STTService):
 
                 if self._final_transcription_buffer or non_final_transcription:
                     final_text = "".join(
-                        map(lambda token: token["text"], self._final_transcription_buffer)
+                        map(
+                            lambda token: token["text"],
+                            self._final_transcription_buffer,
+                        )
                     )
                     non_final_text = "".join(
                         map(lambda token: token["text"], non_final_transcription)