pipecat-ai

Added

• Added on_user_turn_message_added event handler on LLMUserAggregator, with a new UserTurnMessageAddedMessage arg type. It fires when the user aggregator writes a message to the LLM context, carrying the finalized turn text. In cascade mode it coincides with on_user_turn_stopped; in realtime mode (when realtime_service_mode=True on the aggregator pair) it's the canonical way to subscribe to "context just updated, here's the user text" (since the on_user_turn_stopped event fires before the message is finalized, with UserTurnStoppedMessage.content=None). Note that there's been no change to on_assistant_turn_stopped.
  (PR #4533)

• Added RealtimeServiceMetadataFrame, broadcast at pipeline start by realtime LLM services (OpenAI Realtime, Azure Realtime, Inworld, Grok/xAI Realtime, Gemini Live, AWS Nova Sonic, Ultravox). This frame can be used by other processors in the pipeline to configure themselves accordingly. Today, it only advertises two things: that a realtime service is present in the pipeline (indicated by the fact that the frame is sent at all), and emits_user_turn_frames, which says whether the realtime service can emit its own UserStartedSpeakingFrame and UserStoppedSpeakingFrames (suggesting local VAD/turn detection may not be needed in the pipeline).
  (PR #4533)

• Added to our examples "locally-driven-turns" variants for:

  • OpenAI Realtime (realtime-openai-locally-driven-turns.py)
  • Grok Realtime (realtime-grok-locally-driven-turns.py)
  • Inworld Realtime (realtime-inworld-locally-driven-turns.py)

  These join realtime-gemini-live-locally-driven-turns.py in showing how to configure each realtime service so that its turn-taking is dictated by local turn detection (e.g. VAD + smart turn analyzer).
  (PR #4533)

• Added a startup WARNING log on realtime LLM services that don't emit UserStartedSpeakingFrame/UserStoppedSpeakingFrame (Gemini Live, AWS Nova Sonic, Ultravox). The log is meant to draw attention to a couple of things:

  • That other processors in the pipeline (e.g. RTVI) may expect turn frames, and that the developer can enable local VAD/turn detection to supply them, and, relatedly
  • That when using local turn detection, local turns may NOT perfectly align with the "ground truth" of server-decided turns, so they should be thought of as APPROXIMATE (unless local turn detection is driving the realtime service's turns, in which case there's no separate server-decided ground truth)

  (The warning also serves as a little nudge to the realtime service providers: providing a "ground truth" signal of when the provider thinks the user has started or stopped speaking is very helpful to app developers!)
  (PR #4533)

• Added a realtime_service_mode: bool kwarg on LLMContextAggregatorPair, for opting into a set of behaviors tailored for use with realtime (speech-to-speech) services. Setting realtime_service_mode=True does three things: 1. Decouples context writes from the UserStoppedSpeakingFrame signal. Instead, the assistant response start triggers the user message writes. This ensures that context is written properly even when the realtime service provides no turn frames and local turn detection (i.e. local VAD) is disabled. This mechanism also enables the next point. 2. Lets UserStoppedSpeakingFrame fire without waiting for transcripts. When local turn detection is configured to drive realtime service conversations, UserStoppedSpeakingFrame is the signal that triggers assistant responses. By letting this frame fire earlier, we reduce latency. 3. Replaces the default turn strategies with ExternalUserTurnStartStrategy and ExternalUserTurnStopStrategy when the realtime service advertises that it emits its own turn frames. Various realtime services (OpenAI Realtime, Azure, Grok, Inworld) emit their own turn frames; in that case the External strategies fire on_user_turn_started / on_user_turn_stopped from the server-emitted UserStartedSpeakingFrame / UserStoppedSpeakingFrame. For realtime services that don't emit those frames — either because they never do (Gemini Live, Nova Sonic, Ultravox) or because server-side turn detection has been disabled at runtime (e.g. OpenAI Realtime with turn_detection=False, in locally-driven-turns setups) — the defaults stay in place so locally-driven turn detection (e.g. local VAD) can fire the events. Passing custom user_turn_strategies opts out of the swap.

  Note that when realtime_service_mode=True, you should listen for the new on_user_turn_message_added event to get the newly-added user message rather than on_user_turn_stopped, which no longer carries it.
  (PR #4533)

• Added private_endpoint parameter to AzureTTSService and AzureHttpTTSService for connecting via Private Link or custom domain endpoints, matching existing AzureSTTService support.
  (PR #4549)

• Added will_be_spoken field to AggregatedTextFrame. Set to True by the TTS service just before synthesis, allowing downstream processors and observers to know whether TTS will speak a given text segment before audio begins.
  (PR #4559)

• Added AggregatedTextProgressFrame — a new frame emitted alongside each TTSTextFrame during word-timestamp playback. It carries accumulated_text (text already spoken) and remaining_text (text not yet spoken) for the active segment, enabling downstream consumers such as the RTVI observer to do word-level highlighting without coupling to internal sequencer state.
  (PR #4559)

• Added AICQuailVADAnalyzer (pipecat.audio.vad.aic_quail_vad), a noise-robustVoice Activity Detection analyzer powered by the standalone Quail VAD 2.0 model from the ai-coustics SDK (aic-sdk~=2.3.0). It owns its own Processor and works independently of AICFilter, so it can sit before or after enhancement in the pipeline. Defaults to the published quail-vad-2.0-xxs-16khz model; supply model_id/model_path to override.
  (PR #4588)

• Added continuous_partials and interruption_delay connection parameters to the AssemblyAI streaming STT service (u3-rt-pro only). continuous_partials defaults to True so voice agents receive interim transcripts at a steady cadence during long turns; interruption_delay (0–1000 ms) overrides how soon the first partial is emitted. Both are exposed via AssemblyAISTTService.Settings and are omitted for non-u3-rt-pro models.
  (PR #4593)

• Added a user_audio_preroll_secs parameter to GeminiLiveLLMService controlling how much "pre-roll" audio is replayed (sent to Gemini Live) when the user turn start is confirmed, in locally-driven-turns mode (server-side VAD disabled). Defaults to None, auto-sizing the pre-roll duration from the upstream VAD's start_secs (which assumes VAD drives turn starts); set it explicitly when using a non-VAD turn-start strategy.
  (PR #4597)

• Added a user_audio_preroll_secs parameter to OpenAIRealtimeLLMService controlling how much "pre-roll" audio is replayed (re-appended to the input audio buffer) when the user turn start is confirmed, in locally-driven-turns mode (server-side turn detection disabled). Defaults to None, auto-sizing the pre-roll duration from the upstream VAD's start_secs (which assumes VAD drives turn starts); set it explicitly when using a non-VAD turn-start strategy.
  (PR #4599)

• Added word-level timestamp support to SmallestTTSService. Enabled by default via the word_timestamps constructor argument, it emits per-word TTSTextFrames aligned to audio playback so downstream consumers (captions, lip-sync, RTVI) receive word timing. Timestamps from each TTS request are offset onto the turn's continuous playback timeline, so multi-sentence turns stay correctly ordered. Available on Smallest's word-timestamp-capable voices; other voices simply emit no word events, so leaving it on is safe. Pass word_timestamps=False to fall back to whole-text frames.
  (PR #4612)

• Added a profanity setting to AzureSTTService (via settings=AzureSTTService.Settings(profanity=...)) controlling how Azure handles profanity in transcripts. Accepts "raw" (no masking), "masked" (Azure default, replaces profane words with ****), or "removed" (drops profane words). Defaults to None (keeps the Azure SDK default of "masked"). Use "raw" for non-English deployments where Azure's profanity list over-eagerly masks ordinary words. The setting is runtime-updatable and triggers a reconnect when changed.
  (PR #4620)

• WhatsApp connection_callback now receives the full call metadata (WhatsAppConnectCall) as a second argument, available in bot code via runner_args.body. This gives bots access to the caller's phone number, call ID, direction, and timestamp without any extra API calls.
  (PR #4622)

  Added the pipecat create project-scaffolding CLI to pipecat-ai, available via the optional cli extra. Install it with uv tool install "pipecat-ai[cli]" (add --with pipecatcloud to enable pipecat cloud), then run pipecat create to scaffold a new bot project. The CLI dependencies are optional, so they are not pulled into a plain pip install pipecat-ai.
  (PR #4631)

  pipecat create takes an optional target directory: pass a path — for example pipecat create . — to scaffold directly into that directory (the same convention as npm create vite@latest .), or omit it to nest the project under a <project-name>/ subfolder. The project name defaults to the target directory's basename, and --name overrides it.
  (PR #4631)

• Added websocket to the development runner's -t/--transport choices, so you can now run python bot.py -t websocket to restrict the server to the plain WebSocket transport (served at /ws-client). The startup banner prints a websocket-specific message with the prebuilt UI and ws(s)://host:port/ws-client endpoint.
  (PR #4636)

• Direct functions advertised in an LLMContext are now registered automatically — no separate registration call. List a direct function in LLMContext(tools=[...]), or push an LLMSetToolsFrame to change tools mid-session, and its handler is registered. The advertised tool set is the single source of truth: dropping a direct function unregisters its handler too. Also applies across LLMSwitcher member LLMs.
  (PR #4654)

• LLMContext(tools=...) and LLMSetToolsFrame now accept a plain list of direct functions and/or FunctionSchema objects, not just a ToolsSchema.
  (PR #4654)

• Added an optional @tool_options(cancel_on_interruption=..., timeout_secs=...) decorator for overriding a direct function's call options; defaults apply otherwise.
  (PR #4654)

• Added PipelineFlushFrame, a control frame for draining the pipeline. Push it downstream and the pipeline worker bounces it back upstream so it round-trips through every processor, then sets its event. Await that event to know all in-flight frames queued ahead of the probe have been processed (e.g. to let the pipeline settle after an interruption before injecting a new frame). It's an UninterruptibleFrame, so the probe survives an InterruptionFrame and still completes its round-trip.
  (PR #4655)

• Added pipecat.evals, a behavioral eval framework for Pipecat bots, usable both as a library and from the CLI. A YAML scenario describes a scripted conversation and the semantic events expected back from the bot (transcriptions, LLM/TTS responses, function calls) with optional latency budgets and natural-language criteria judged by an LLM, in text or audio mode (audio synthesizes the user's speech and transcribes the bot's actual audio). In code, EvalScenario.load() parses a scenario and EvalSession.from_scenario(...).run() runs it against a bot, returning a structured EvalResult (with EvalManifest.load() and EvalSuite.run() for the multi-bot path). The new pipecat eval run (against an already-running bot) and pipecat eval suite (a manifest mapping bots to the scenarios they run) commands wrap the same library and are also reachable as python -m pipecat.evals. Bots opt in by exposing the -t eval transport.
  (PR #4655)

• Added a bot-interrupted RTVI server message, emitted when the bot's in-flight output is cut off (a VAD-detected user barge-in or a programmatic interrupt), so clients can drop whatever the bot was mid-saying.
  (PR #4655)

• Added an opt-in --eval flag to pipecat create (and an Enable evals? wizard prompt, off by default) that makes the generated bot eval-ready without any manual edit:

  • an "eval" entry in the bot's transport_params, so the bot is runnable with -t eval. The entry mirrors the bot's audio/video settings and is inert unless the bot is run with -t eval.
  • runnable starter scenarios in server/evals/ that pass against the freshly scaffolded bot and double as schema references to copy when adding more: starter_text.yaml (text mode, the fast inner loop; cascade bots only) and starter_audio.yaml (the full audio round trip, the only mode for realtime speech-to-speech bots).
  • the dependencies to run them from the project's own environment: the cli extra (the pipecat eval command) plus kokoro and moonshine (the harness's local speech stack), so audio-mode evals run with no extra setup and no API keys.
    (PR #4664)

• Added filter_repeated_sequences parameter to MarkdownTextFilter.InputParams to allow disabling repeated sequence removal.
  (PR #4674)

• Added support for Belgium german in transcription languages
  (PR #4682)

• Added MoonshineSTTService, a local speech-to-text service backed by Moonshine. It runs a small, fast ASR model on the CPU via ONNX Runtime, so it needs no GPU and no API key (the model downloads once on first use and is cached). Install with pip install "pipecat-ai[moonshine]" and choose the model via MoonshineSTTService.Settings(model=...) (a Model enum member or string): Model.TINY, Model.BASE, or a streaming model run in batch (Model.TINY_STREAMING, Model.SMALL_STREAMING (default), Model.MEDIUM_STREAMING). See examples/voice/voice-moonshine.py.
  (PR #4683)

• New features for the Vonage WebRTC transport

  • Captions support
  • Individual audio stream subscription support
  • Updated to Vonage Video Connector library v1.0.2
    (PR #4686)

• FunctionSchema now accepts an optional handler. When set, the LLM service registers it automatically wherever the schema is advertised in an LLMContext (or via an LLMSetToolsFrame), so no separate register_function call is needed. This extends the existing auto-registration of direct functions to FunctionSchema-based tools: the advertised tool set stays the single source of truth, so dropping a handler-carrying schema unregisters its handler too. A FunctionSchema without a handler stays advertise-only. Decorate the handler with @tool_options to override its default call options (cancel_on_interruption, timeout_secs), the same decorator direct functions use.
  (PR #4709)

• Added pipecat init, which makes a project agent-ready by writing a Pipecat coding-agent guide (AGENTS.md plus a CLAUDE.md that imports it) and developer guidance (GETTING_STARTED.md — MCP setup, how to write a good first prompt with a copyable example, what to expect from the session) into the project, so an AI coding assistant picks up Pipecat conventions automatically and then scaffolds the app with pipecat create. Run pipecat init (prompts for a directory), pipecat init my-bot, or pipecat init .; re-running refreshes AGENTS.md while preserving an existing CLAUDE.md (pass --force to overwrite it). The written AGENTS.md ends with a provenance footer naming the pipecat-ai version that wrote it, so a stale guide is detectable and refreshable.
  (PR #4710)

• Added context carryover support to AssemblyAISTTService for Universal-3 Pro streaming (u3-rt-pro). A new agent_context setting seeds the agent's most recent reply at connect time, and AssemblyAISTTService.update_agent_context() updates it mid-session via an UpdateConfiguration message (no reconnect). Giving the model the agent's last reply improves transcription of the user's next turn — short answers, spelled-out entities, and similar-sounding words. A previous_context_n_turns setting controls how many prior entries are carried forward (set to 0 to disable carryover entirely). U3 Pro features are recognized for the whole u3-rt-pro family, including the u3-rt-pro-beta-1 variant.

  • Added universal-3-5-pro as a supported AssemblyAISTTService model. It is recognized as part of the Universal-3 Pro family, so every u3-rt-pro feature (built-in turn detection, prompting, continuous partials, interruption_delay, context carryover, and voice focus) applies to it as well.

  • Added voice_focus and voice_focus_threshold settings to AssemblyAISTTService (Universal-3 Pro models). Set voice_focus to "near-field" or "far-field" to isolate the primary voice and suppress background noise; voice_focus_threshold (0.0–1.0) tunes how aggressively background audio is suppressed.
    (PR #4712)

• Added the "Add a WebRTC transport for local testing?" option to the Daily PSTN and Twilio + Daily SIP scenarios in pipecat init, so the generated bots can also be run locally with the SmallWebRTC or Daily client.
  (PR #4715)

• Realtime and speech-to-speech LLM services that take tools at construction now accept a plain list of standard tools (direct functions and/or FunctionSchema objects), not just a ToolsSchema — matching LLMContext(tools=...). Applies to GeminiLiveLLMService / GeminiLiveVertexLLMService (tools=), AWSNovaSonicLLMService (tools=), UltravoxRealtimeLLMService (one_shot_selected_tools=), and session_properties.tools on the OpenAIRealtimeLLMService / AzureRealtimeLLMService / GrokRealtimeLLMService / InworldRealtimeLLMService.
  (PR #4758)

• Added STTService.process_assistant_turn(text) hook that subclasses can override to feed the completed bot reply to a provider-side context carryover API. The base implementation is a no-op; STTService now handles LLMContextAssistantTurnFrame and calls this method automatically.
  (PR #4759)

• Added LLMContextAssistantTurnFrame, broadcast by LLMAssistantAggregator when a bot turn completes, carrying the aggregated reply text and start timestamp.
  (PR #4759)

• Added endpoint_sensitivity to SonioxSTTService.Settings, a float in [-1.0, 1.0] that controls how aggressively Soniox emits speech endpoints. Higher values finalize turns sooner; lower values delay them. Introduced in the Soniox v5 model; earlier models reject it.
  (PR #4772)

• DailyTransport can now publish a screenAudio output track, mirroring screenVideo. Add "screenAudio" to DailyParams.audio_out_destinations (and optionally configure it via custom_audio_track_params["screenAudio"]), then write audio frames with transport_destination="screenAudio". Requires daily-python>=0.29.0.
  (PR #4775)

• Added an evals extra that bundles the pipecat eval command (the cli extra) with the harness's default local, no-API-key models: Kokoro (user-turn TTS) and Moonshine (bot-speech transcription). Install pipecat-ai[evals] so uv run pipecat eval run works out of the box. Scaffolded projects (pipecat init) that enable evals now depend on pipecat-ai[evals].
  (PR #4776)

• RTVIObserver can now emit raw VAD user speaking events (vad-user-started-speaking / vad-user-stopped-speaking), driven directly by the VAD signal and independent of turn finalization (unlike user-started-speaking / user-stopped-speaking, which a turn strategy may gate or defer). Enable with RTVIObserverParams(vad_user_speaking_enabled=True) (off by default), or at runtime via RTVIConfigureObserverFrame.
  (PR #4785)

Changed

• Migrated all realtime LLM service examples (OpenAI Realtime, Azure Realtime, Inworld, Grok/xAI Realtime, Gemini Live, Gemini Live Vertex, AWS Nova Sonic, Ultravox) to use LLMContextAggregatorPair(..., realtime_service_mode=True). Where examples previously wired SileroVADAnalyzer into LLMUserAggregatorParams as a workaround for missing turn frames, the local VAD has been removed; LLMContextAggregatorPair's realtime_service_mode makes this safe in terms of context-writing. Transcript-logging user-side event handlers have moved from on_user_turn_stopped to the new on_user_turn_message_added event, which carries the finalized message text (the turn-stopped event fires before the message is finalized in realtime service mode). Examples for services without server-side user-turn frames (Gemini Live, AWS Nova Sonic, Ultravox) include a comment block explaining how to add local VAD if needed. Each base example now also subscribes to on_user_turn_stopped — active for services that emit server-side user-turn frames (OpenAI Realtime, Azure Realtime, Grok, Inworld) and commented-out for those that don't (with the same opt-in path as the local-VAD block).
  (PR #4533)

• UserTurnStoppedMessage.content is now typed str | None. In realtime mode (realtime_service_mode=True on LLMContextAggregatorPair) the user message isn't finalized at turn-stop time, so content is None; subscribers wanting the finalized text should use the new on_user_turn_message_added event. Behavior in cascade (STT -> LLM -> STT) pipelines is unchanged.
  (PR #4533)

• SpeechTimeoutUserTurnStopStrategy, TurnAnalyzerUserTurnStopStrategy, and ExternalUserTurnStopStrategy now accept a wait_for_transcript: bool = True kwarg. When flipped to False, the strategy signals end-of-turn as soon as its requirements are met, minus waiting for transcripts — useful when you intend to configure local turn detection to drive realtime service conversations, where waiting for transcripts is unnecessary latency. LLMContextAggregatorPair flips this for you when realtime_service_mode=True.
  (PR #4533)

• Updated Smallest AI TTS plugin for Waves v4.0.0 API:

  • New WebSocket endpoint /waves/v1/tts/live (previously /waves/v1/{model}/get_speech/stream)
  • Model is now sent in each message payload instead of the URL, eliminating reconnection on model change
  • Updated model names: lightning_v3.1 and lightning_v3.1_pro (underscore convention)
  • Added output_format setting supporting pcm, mp3, wav, ulaw, alaw
  • Default model changed to lightning_v3.1_pro (with meher as its default voice)
  • Breaking: SmallestTTSModel.LIGHTNING_V2 removed; consistency, similarity, enhancement settings removed
    (PR #4535)

• ⚠️ RTVI protocol version bumped to 2.0.0. The bot-output message now includes will_be_spoken, spoken_status ("new" / "in-progress" / "completed"), spoken_progress (accumulated/remaining text), and segment_id fields. Clients on any 1.x protocol are still served with the legacy format; all other pre-2.x clients are rejected.
  (PR #4559)

• bot_output_transforms now supports a 4-parameter progress-aware signature: (text, agg_type, accumulated_text, remaining_text) -> BotOutputTransformResult. When called for a progress event, accumulated_text and remaining_text are populated and the transform must return a BotOutputTransformResult with those fields set, enabling word-level transforms on the client side.
  (PR #4559)

• Updated aic-sdk dependency to ~=2.3.0. The AIC_SDK_LICENSE environment variable replaces the previous AIC_LICENSE_KEY so the variable matches the SDK's canonical name; users must update their .env files.
  (PR #4588)

• Aligned the deprecation docstrings in LLMUserAggregatorParams with the project's documented convention by removing redundant inline [DEPRECATED] tags, keeping only the .. deprecated:: Sphinx directive.
  (PR #4592)

• AzureSTTService now marks final transcripts as finalized. Azure's RecognizedSpeech event is by definition the final recognition for an utterance, so the emitted TranscriptionFrame carries finalized=True. This lets downstream user-turn stop strategies (e.g. SpeechTimeoutUserTurnStop) take their finalized fast-path instead of waiting for VAD events that may never arrive on short replies.
  (PR #4620)

• ⚠️ The mem0 extra now requires mem0ai>=2,<3. Mem0MemoryService was updated for the mem0 2.0.0 breaking changes: entity IDs (user_id/agent_id/run_id) are now passed via filters= to the local client (top-level kwargs raise ValueError in mem0 2.x), and the removed version/output_format parameters are no longer sent to the cloud client. Note that mem0 2.0.0 also flips the rerank default from True to False and makes add() async server-side (stored memories are queryable once processed).
  (PR #4626)

• GradiumSTTService now defaults delay_in_frames to 12 (960ms) instead of leaving it unset (which used the server default of 10/800ms). The higher default allows more context for improved transcription accuracy. Set delay_in_frames explicitly to 7-8 for faster responses.
  (PR #4632)

• GradiumSTTService has an updated ttfs_p99_latency value of 0.62 seconds.
  (PR #4632)

• Bumped pipecat-ai-prebuilt to 1.0.2 in the runner extra, updating the prebuilt client UI served by the development runner.
  (PR #4634)

• ⚠️ Changed the default of TTSSpeakFrame.append_to_context from None to True. The old None behavior was situation-dependent and hard to reason about: the spoken text always reached the assistant aggregator's buffer, but whether it was committed to the LLM context depended on what surrounded the frame — committed when the frame was inside an assistant response or immediately followed by one, but silently discarded when it was standalone and followed by a user turn (the interruption cleared the buffer before anything flushed it). True is a predictable default: programmatically-spoken text is recorded in the context unless you opt out with append_to_context=False. BusTTSSpeakMessage.append_to_context now defaults to True to match.
  (PR #4642)

• Switched the aws extra from aioboto3 to aiobotocore. Pipecat only uses the low-level client API, and aiobotocore is the async library that aioboto3 wraps, so depending on it directly drops an unnecessary wrapper layer. AWS service initialization now uses aiobotocore.session.get_session() and session.create_client(...); public APIs and credential resolution are unchanged.
  (PR #4643)

• websockets is now a core dependency of pipecat-ai instead of the websockets-base optional extra. The websockets-base extra has been removed; service extras that used to pull it in (Cartesia, Deepgram, ElevenLabs, OpenAI, Google, and others) still work unchanged, and websockets is now always installed. If you previously installed pipecat-ai[websockets-base] directly, just drop the extra since pip install pipecat-ai now includes it.
  (PR #4658)

• Renamed the @tool decorator's timeout argument to timeout_secs, matching register_function(). timeout still works as a deprecated alias and will be removed in a future version.
  (PR #4671)

• WhisperSTTService's Model and MLXModel are now StrEnum, so a member is the string itself (e.g. Model.TINY == "tiny"). Passing a Model/MLXModel member or a plain string both keep working.
  (PR #4684)

• Bumped the daily extra's daily-python dependency to >=0.29.1,<1.
  (PR #4685)

• LLMWorker now enables the worker's automatic RTVI support when it is not bridged (bridged=None), so a standalone LLMWorker driving its own transport gets the RTVIProcessor/RTVIObserver pair like any PipelineWorker. Bridged child workers keep RTVI disabled, since the transport worker owns the client-facing RTVI machinery.
  (PR #4690)

• Removed the asyncio.sleep(0) workarounds that let a just-created timer task start before a possible immediate cancellation. TaskManager.create_task() now cleans up never-started coroutines centrally, so the yields served no purpose.
  (PR #4692)

• Worker frames (e.g. EndWorkerFrame) should now be pushed downstream with a plain push_frame(frame), so frames queued ahead of them are flushed before the worker acts on them. Pushing them upstream still works.
  (PR #4705)

• register_function now reads a handler's call options (cancel_on_interruption, timeout_secs) from its @tool_options decorator when they aren't passed explicitly, matching how direct functions resolve them (explicit argument > @tool_options > default). Previously the decorator was ignored on this path.
  (PR #4709)

• BaseLLMAdapter.from_standard_tools now raises UserWarning instead of DeprecationWarning when built-in tools can't be injected because the supplied tools aren't a ToolsSchema — it advises about the tools format and is not a deprecation.
  (PR #4726)

• Deprecated classes and functions are now marked with the PEP 702 @deprecated decorator, so type checkers and IDEs (pyright/Pylance reportDeprecated, mypy's deprecated error code) flag and strike through deprecated usages statically. Several deprecated classes that previously emitted no runtime warning now raise DeprecationWarning when used, and deprecation messages now state a concrete removal version (e.g. 2.0.0) instead of "a future release".
  (PR #4726)

• pipecat create now infers --bot-type from the chosen transports in non-interactive mode, so the flag is optional: a bot is telephony when any transport is a telephony transport (twilio, telnyx, plivo, exotel, daily_pstn, twilio_daily_sip) and web otherwise. Pass --bot-type explicitly to override (it's still validated and cross-checked against the transports); the interactive wizard is unchanged.
  (PR #4735)

• Realtime LLM services now auto-register the handlers bundled on the tools passed at construction time, so a separate register_function() call is no longer needed — matching how context-advertised tools (a direct function, or a FunctionSchema with its handler set) already register. Applies to GeminiLiveLLMService / GeminiLiveVertexLLMService (tools=), UltravoxRealtimeLLMService (one_shot_selected_tools=), AWSNovaSonicLLMService (tools=), and OpenAIRealtimeLLMService / AzureRealtimeLLMService / GrokRealtimeLLMService / InworldRealtimeLLMService (session_properties.tools).
  (PR #4758)

• Updated SonioxSTTService default model from stt-rt-v4 to stt-rt-v5.
  (PR #4772)

• The Kokoro TTS model cache moved to ~/.cache/pipecat/kokoro-onnx (previously ~/.cache/kokoro-onnx), so Pipecat's cached files live under a single namespaced directory.
  (PR #4776)

Deprecated

• Deprecated the 2-parameter bot_output_transforms signature (text, agg_type) -> str. Transforms using it will still work but emit a DeprecationWarning at registration time. Update to the 4-parameter signature (text, agg_type, accumulated_text, remaining_text) -> BotOutputTransformResult to support word-level progress transforms.
  (PR #4559)

• ⚠️ Deprecated AICVADAnalyzer (pipecat.audio.vad.aic_vad) and AICFilter.create_vad_analyzer(). Both are tied to AICFilter's model-internal VAD path. Use AICQuailVADAnalyzer instead — the standalone Quail VAD 2.0 model is the noise-robust VAD differentiator going forward. Both surfaces will be removed in Pipecat 1.6.0 (breaking change shipped in a minor release, per maintainer guidance for plugins).
  (PR #4588)

• The single-argument connection_callback(connection) signature for WhatsAppClient.handle_webhook_request is deprecated. Update callbacks to accept (connection, call: WhatsAppConnectCall) to receive call metadata alongside the WebRTC connection. The old signature still works but emits a DeprecationWarning.
  (PR #4622)

• Deprecated Mem0MemoryService.InputParams.api_version. It is no longer used — mem0 2.0.0 removed the api_version/output_format parameters from the client. Setting it now emits a DeprecationWarning.
  (PR #4626)

• Deprecated passing append_to_context=None to TTSSpeakFrame (and BusTTSSpeakMessage). None is no longer a supported value: it is coerced to True with a warning and will be unsupported in a future release. Pass True or False explicitly. See the corresponding "Changed" entry for the full rationale behind the new True default.
  (PR #4642)

• Deprecated LLMService.register_direct_function() / unregister_direct_function() and LLMSwitcher.register_direct_function(). Advertise direct functions in LLMContext(tools=[...]) or via an LLMSetToolsFrame instead — handlers are registered and unregistered automatically. These will be removed in a future version.
  (PR #4671)

• ⚠️ Deprecated TaskFrame, TaskSystemFrame, EndTaskFrame, StopTaskFrame, CancelTaskFrame and InterruptionTaskFrame. Use WorkerFrame, WorkerSystemFrame, EndWorkerFrame, StopWorkerFrame, CancelWorkerFrame and InterruptionWorkerFrame instead, matching the PipelineWorker naming. The old names remain as isinstance-compatible aliases that emit a DeprecationWarning on construction.
  (PR #4705)

• Renamed WebsocketServerTransport to SingleClientWebsocketServerTransport to make it explicit that the server handles a single client at a time. The supporting WebsocketServerParams, WebsocketServerCallbacks, WebsocketServerInputTransport, and WebsocketServerOutputTransport classes were renamed with the same SingleClient prefix. The old names remain as deprecated aliases and will be removed in 2.0.0.
  (PR #4774)

Fixed

• Fixed output image resizing for generated images when video output dimensions differ from the source image size by consistently using Pillow pixel modes instead of encoded formats.
  (PR #4483)

• Fixed a benign ERROR log line emitted by UltravoxRealtimeLLMService during client-driven teardown. Adds an exception catch which guards the disconnecting case.
  (PR #4519)

• Fixed InworldRealtimeLLMService not supporting manual-mode turn detection (session_properties.audio.input.turn_detection=None). Previously _handle_user_stopped_speaking and _handle_interruption assumed Inworld's server-side VAD handled commit/cancel/response.create automatically and were no-ops on the client side. In manual mode the server doesn't, so local-VAD-driven turns stalled: the bot never responded after the user stopped speaking, and interruptions didn't cancel the in-flight response. Wire the explicit InputAudioBufferCommitEvent + ResponseCreateEvent on user-stopped-speaking and InputAudioBufferClearEvent + ResponseCancelEvent on interruption, gated on a new _is_manual_turn_detection() check (mirroring the pattern in OpenAIRealtimeLLMService).
  (PR #4533)

• InworldRealtimeLLMService and GrokRealtimeLLMService no longer broadcast UserStartedSpeakingFrame/UserStoppedSpeakingFrame when configured for manual (locally-driven) turn detection. Both services' server-side speech-started/stopped events fire in manual mode too, but in that setup turn frames are expected to come from local turn detection (e.g. a vad_analyzer in LLMUserAggregatorParams) — without the gate, the services were broadcasting alongside the locally-emitted frames, producing duplicate on_user_turn_* events. OpenAI Realtime was already correct here: its server doesn't fire speech events in manual mode at all.
  (PR #4533)

• Fixed Ultravox Realtime not surfacing server-side interruption. The server sends a playback_clear_buffer message when the user interrupts the bot mid-speech, instructing clients to drop buffered output audio; this was previously unhandled, so BaseOutputTransport kept playing the buffered audio and the bot kept talking past the interruption. Ultravox now broadcasts InterruptionFrame on playback_clear_buffer. This was previously masked by enabling local VAD on the user aggregator, which generated UserStartedSpeakingFrame and triggered the aggregator-side interruption path; the fix makes the behavior correct without local VAD as a workaround.
  (PR #4533)

• Fixed GrokRealtimeLLMService stalling the conversation when Grok returns an error in response to a response.cancel event sent while no response is active on the server. This happens routinely in manual-turn-detection mode: when the user starts speaking after the bot has finished, Pipecat broadcasts an InterruptionFrame and the service sends ResponseCancelEvent, which Grok rejects with "Cancellation failed: no active response found". The existing error-suppression list only matched OpenAI's response_cancel_not_active / conversation_already_has_active_response error codes, but Grok uses different codes for the same conditions — so the error fell through to the fatal-error path and exited the WebSocket receive loop, preventing any further server events from being processed. The suppression now also matches on the error message substring ("no active response", "already has an active response"), so these benign races get logged at debug and the receive loop keeps running.
  (PR #4533)

• Fixed AWS Nova Sonic not surfacing server-side interruption. When the user interrupted the bot mid-response, the INTERRUPTED stop reason was acknowledged internally but no InterruptionFrame was emitted, so BaseOutputTransport kept draining its audio buffer and the bot kept talking past the interruption. Nova Sonic now broadcasts InterruptionFrame on both INTERRUPTED paths (text-stage and audio-stage). This was previously masked by enabling local VAD on the user aggregator, which generated UserStartedSpeakingFrame and triggered the aggregator-side interruption path; the fix makes the behavior correct without local VAD as a workaround.
  (PR #4533)

• Fixed pipeline shutdown hanging on LiveKit when the remote peer disconnected mid-stream. The trailing audio_out_end_silence_secs write is now bounded by a timeout.
  (PR #4578)

• Fixed the start of user speech being clipped from transcripts when GeminiLiveLLMService is configured for locally-driven turns (server-side VAD disabled). The problem was that any audio sent up to Gemini Live before sending activity_start (sent when user turn start is confirmed) seemed to get discarded; the service now replays (sends to Gemini Live) a short audio "pre-roll" right after activity_start, so the onset is preserved.
  (PR #4597)

• Fixed the start of user speech being clipped from transcripts when OpenAIRealtimeLLMService is configured for locally-driven turns (server-side turn detection disabled). The problem was that the speech onset already sent to OpenAI got discarded when the service cleared its input audio buffer on barge-in (which it does when the user turn start is confirmed); the service now replays (re-appends to the input audio buffer) a short audio "pre-roll" right after the clear, so the onset is preserved.
  (PR #4599)

• 422 validation errors now log the full error details and raw request body for all transports (WhatsApp, WebRTC, telephony, etc.), making malformed payloads easier to debug. Previously this logging only applied to WhatsApp routes.
  (PR #4622)

• Fixed InworldTTSService logging a spurious "no websocket connected, will try to reconnect" warning and firing a redundant second reconnect when the initial connection attempt failed. The service now returns an ErrorFrame immediately if the websocket is unavailable after _connect(), matching the behaviour of ElevenLabsTTSService.
  (PR #4635)

• Fixed SarvamTTSService (WebSocket) emitting BotStoppedSpeakingFrame late. The service never produced a TTSStoppedFrame on synthesis completion, so end-of-turn was detected only by the stop_frame_timeout_s idle timer, causing BotStoppedSpeakingFrame to lag the actual end of audio by up to that timeout (especially for short utterances or a raised stop_frame_timeout_s). The service now requests Sarvam's completion event (send_completion_event) and emits TTSStoppedFrame as soon as the final event arrives, so the bot-stopped-speaking event tracks the end of audio. The idle timeout remains as a fallback.
  (PR #4639)

• Fixed a spurious RuntimeWarning: coroutine '...' was never awaited emitted by TaskManager.create_task() when a task is cancelled before its coroutine starts running. The wrapper now closes the un-started coroutine on cancellation, so the warning no longer fires. This surfaced, for example, when combining TurnAnalyzerUserTurnStopStrategy with another stop strategy that force-completes the turn (cancelling the analyzer's timeout task before it ran), and when a function call is cancelled by a user-turn-start interruption race (the LLMService._run_function_call warning, #4339). A local await asyncio.sleep(0) workaround in _run_function_call that existed only to dodge this warning has been removed now that it is handled centrally. The turn/cancellation behavior was already correct; only the noisy warning is removed.
  (PR #4644)

• Fixed LiveKitTransport leaking audio/video stream readers when a track is unsubscribed: the owned rtc.AudioStream/rtc.VideoStream and its producer task are now closed and cancelled on unsubscribe (and on a re-subscribe for the same participant), so a client republishing its mic (e.g. mute/unmute or text↔voice toggles) no longer accumulates concurrent producers that interleave audio into the shared queue and silence downstream STT.
  (PR #4650)

• Fixed TTSService emitting a second LLMFullResponseEndFrame (with a new id) at the end of an audio context when push_text_frames is False, which caused RTVIObserver to send a duplicate bot-llm-stopped message per LLM response. The original end frame received in process_frame is now held per context_id and re-pushed, preserving its id.
  (PR #4653)

• Fixed a frame-ordering race in bridged workers: frames received from the WorkerBus were pushed directly into the pipeline from the bus edge, so they could interleave with (or overtake) frames the worker had queued itself via queue_frame()/queue_frames(). A bus inbound frame (e.g. an LLMContextFrame from a concurrent user input) could reach the LLM in the middle of a multi-frame update such as a flow's set_node (LLMMessagesUpdateFrame + LLMSetToolsFrame), generating against the previous node's context. Bus inbound frames are now serialized through the worker's frame queue, so both paths share one FIFO.
  (PR #4656)

• Fixed interruption handling for standalone TTSSpeakFrame(append_to_context=True) utterances (those not part of an LLM response). Previously, when the user interrupted such an utterance:

  • on_assistant_turn_stopped didn't fire
  • partially-spoken text wasn't recorded to the context (for TTS services that support word timestamps)

  The problem was that these utterances have no LLMFullResponseStartFrame to open the assistant turn, so there was no open turn for the interruption to stop. The assistant aggregator now uses a new TTSStartedFrame.append_to_context to open the turn when the utterance begins.

  As a result of this fix, on_assistant_turn_started timing is improved for standalone TTSSpeakFrame utterances: the event now fires at the start rather than at the end.
  (PR #4665)

• Fixed OpenAIResponsesHttpLLMService raising 'NoneType' object has no attribute 'cached_tokens' on every turn when used with a custom base_url pointing at a third-party Responses API server that omits the OpenAI-specific input_tokens_details / output_tokens_details sub-objects. Token usage parsing now tolerates any field the server omits — the SDK's lenient streaming decoder leaves these as None whether it's a top-level count (input_tokens / output_tokens / total_tokens), a missing detail sub-object, or a missing field inside one — and falls back to 0 in each case, matching the WebSocket OpenAIResponsesLLMService variant.
  (PR #4667)

• Fixed importing pipecat.services.whisper.stt failing on non-macOS platforms when the mlx-whisper extra happened to be installed: mlx_whisper is now only imported on macOS (it's Apple-Silicon only, and elsewhere the package can be installed but unloadable, e.g. a missing libmlx.so). WhisperSTTServiceMLX still imports it lazily when actually used.
  (PR #4684)

• Fixed SambaNovaLLMService failing every completion with its default model: SambaNova Cloud removed Llama-4-Maverick-17B-128E-Instruct, so the default is now Meta-Llama-3.3-70B-Instruct.
  (PR #4687)

• Fixed NebiusLLMService function calls never executing with its default model: Nebius streams openai/gpt-oss-120b tool calls with a broken final fragment (index=1 on a single call), so the default is now Qwen/Qwen3-30B-A3B-Instruct-2507, which streams correctly.
  (PR #4688)

• Fixed a worker-handoff race where activate_worker(deactivate_self=True) left both workers briefly active: the caller's active flag only flipped when its own deactivate message came back over the bus, so the handoff target could activate first and both workers re-broadcast each other's frames (duplicate tool round-trips in the LLM context). The caller now deactivates synchronously before publishing the activate message.
  (PR #4691)

• Fixed AzureTTSService producing no audio when running in a pipeline without an output transport (e.g. headless/offline setups). Audio chunks arrive from the Speech SDK on native threads, and the cross-thread queue hand-off didn't wake an otherwise-idle event loop; the service now marshals those puts onto the loop, so audio is delivered regardless of loop activity.
  (PR #4703)

• Fixed a regression from #4654 where unregistering a tool's handler on its own — via unregister_function / unregister_direct_function, without changing the advertised tool set — was silently undone, because the next LLMContextFrame re-registered the handler from the still-advertised tool (a "zombie"). An explicitly unregistered handler now stays unregistered while its tool remains advertised (so calls hit the missing-handler recovery and the model learns to stop), and is restored only by registering it again, or by re-advertising the tool (removing it from the advertised set, then adding it back).
  (PR #4709)

• Fixed LLMSwitcher.register_direct_function overriding a direct function's @tool_options call options. Its cancel_on_interruption defaulted to True and was forwarded to each member LLM as an explicit value, so a @tool_options(cancel_on_interruption=False) handler was ignored. It now defaults to None and follows the same explicit-arg > @tool_options > default fallback as LLMService.register_direct_function (the fallback was added there in #4654 but not mirrored on the switcher).
  (PR #4709)

• Fixed an audible 200-300 ms gap in audio mixer output (e.g. SoundfileMixer background sound) on every interruption. The output transport now keeps the audio task running and drains the queue instead of cancelling and recreating it when a mixer is active.
  (PR #4714)

• Fixed pipecat init silently ignoring the "Enable evals?" option for the Daily PSTN Dial-out and Twilio + Daily SIP scenarios. Generated bots for these scenarios can now be driven with pipecat eval (-t eval): they fall back to create_transport() when the request body carries no room_url, while the production flow (room and call settings arriving from server.py) is unchanged. In local runs the dial-out and Twilio call-forwarding machinery is skipped and the bot stays silent until spoken to.
  (PR #4715)

• Fixed FastAPIWebsocketTransport stalling pipeline shutdown for ~10s when the client's WebSocket is half-closed (e.g. a telephony call already torn down on the provider's side, leaving the media-streams socket open at the TCP layer but unresponsive). FastAPIWebsocketClient.disconnect() awaited websocket.close() unbounded, so it blocked on the ASGI server's close-handshake timeout, delaying EndFrame propagation and the whole pipeline teardown. The close handshake is now bounded by a new FastAPIWebsocketParams.ws_close_timeout (default 0.5s): the close is still initiated, but disconnect() waits at most that long for the peer's acknowledgment before letting shutdown proceed. Increase ws_close_timeout for high-latency peers that need longer to complete a graceful close.
  (PR #4723)

• Fixed NvidiaLLMService reasoning streams so interrupted or early-cancelled responses clean up correctly and do not leak buffered thought content or leave the wrapped stream open.
  (PR #4743)

• Fixed an issue in AggregatedFrameSequencer where delayed word-timestamps from an interrupted (cleared) TTS context could be emitted as passthrough TTSTextFrames with append_to_context=True, interleaving stale words into the next turn's transcript (observed with Inworld TTS in ASYNC mode). Words for an unknown or cleared context are now dropped instead of corrupting the active turn.
  (PR #4751)

• Fixed RimeTTSService.SPELL() and RimeTTSService.PAUSE_TAG() helpers, which are now static methods. Previously they were defined as instance methods without a self parameter, so calling them on a service instance bound the instance to the first argument and produced incorrect output.
  (PR #4755)

• Fixed SingleClientWebsocketServerTransport (formerly WebsocketServerTransport) so that a new client connection no longer disconnects the client that is already connected. While a client is connected, new connection attempts are now rejected with a warning. The active connection's reference is cleared when the client disconnects or the connection fails, so a new client can connect afterwards.
  (PR #4774)

• Fixed DeepgramSageMakerSTTService raising TypeError on construction. Its default settings still passed vad_events, which was removed from DeepgramSTTService.Settings, so the service (and the voice-deepgram-sagemaker example) crashed on instantiation.
  (PR #4786)

Security

• Added optional HMAC token authentication for WebSocket connections in the development runner. Set PIPECAT_WEBSOCKET_AUTH=token (or pass --ws-auth token) to require clients to call POST /start and obtain a short-lived signed session token before connecting. Tokens are one-time use and expire after 5 minutes.

  • Clients can supply the token via Authorization: Bearer <token> header, ?token=<token> query parameter, or URL path segment (/ws/<token>, /ws-client/<token>) — the path form is recommended for telephony providers like Twilio.
  • Both the telephony WebSocket (/ws) and plain WebSocket (/ws-client) endpoints are protected. Connections with invalid, expired, or replayed tokens are rejected with WebSocket close code 4003.
    (PR #4660)

• Added origin restriction support to WebsocketServerTransport, FastAPIWebsocketTransport, and the development runner to mitigate Cross-Site WebSocket Hijacking (CSWSH). When allowed_origins is configured, connections with a missing or disallowed Origin header are rejected before the WebSocket handshake completes.

  • WebsocketServerParams and FastAPIWebsocketParams gain an allowed_origins: list[str] field. FastAPIWebsocketTransport raises ValueError at construction time if the origin is not allowed.
  • The runner gains --allowed-origins CLI flag and PIPECAT_ALLOWED_ORIGINS environment variable (comma-separated). Both also control the transport params default, so a single env var covers all WebSocket transports uniformly.
  • Default is empty (allow all) — no behaviour change for existing deployments.
    (PR

Note truncated.

• Pipecat pipelines are multi-agent compatible by default. The new multi-agent framework (pipecat.workers) turns every PipelineWorker (previously PipelineTask) into a peer on a shared bus that passes typed messages, dispatches @job work, and coordinates with siblings, while existing single-pipeline code keeps running untouched. examples/multi-worker/ ships ready-to-run patterns: LLM handoff, parallel debate, sidecar code assistants and hardware controllers, distributed deployments over Redis or PGMQ, point-to-point WebSocket proxies, and UI workers driving a web client over RTVI.
  (PR #4493)

• Added UIWorker (pipecat.workers.ui): an LLM worker that observes and drives a client web UI over the RTVI UI channel — for voice agents that act on what the user is looking at. It reads the page's accessibility snapshots, routes client UI events to @ui_event handlers, drives the page with UI commands (scroll_to, highlight, select_text, click, set_input_value), and answers screen-grounded questions. PipelineWorker connects it to the client automatically when RTVI is enabled — no extra wiring.

  • A voice agent delegates a turn via the built-in respond job; the worker returns an answer for the voice LLM to speak, or speaks it verbatim through the agent's TTS with respond_to_job(answer, tts_speak=True).
  • ReplyToolMixin provides a ready-made reply tool (a spoken answer plus the standard UI actions).
  • ui_job_group(...) fans work out to peer workers, surfaced to the client as cancellable progress cards.
  • UI_STATE_PROMPT_GUIDE is drop-in system-prompt text that teaches the LLM the <ui_state> wire format.
    (PR #4540)

• Added VonageVideoConnectorTransport, a new transport integration for real-time Vonage WebRTC sessions using the Vonage Video Connector library.
  (PR #4052)

• Added InceptionLLMService for Inception's Mercury 2 diffusion reasoning model, with support for reasoning_effort and realtime settings.
  (PR #4423)

• Added plain WebSocket transport support to the development runner. Bots can now accept connections from non-telephony WebSocket clients (e.g., browser apps using protobuf framing) via the /ws-client endpoint alongside other transports.
  (PR #4442)

• Added GET /status endpoint to the development runner that reports which transports the running instance accepts (all by default, or the single transport passed via -t).
  (PR #4442)

• Added support for the Rime coda TTS model to RimeTTSService and RimeHttpTTSService. The temperature, top_p, and repetition_penalty settings are not used by coda. Also added a timeScaleFactor setting (for the arcana and coda models) to both services — values above 1.0 slow down audio playback; values below 1.0 speed it up.
  (PR #4511)

• Added max_endpoint_delay_ms to SonioxSTTService.Settings, controlling the maximum delay (500-3000 ms) before endpoint detection finalizes a turn.
  (PR #4521)

• Added LLMService.append_system_instruction(...): append durable text to a service's system instruction so it's included on every inference and survives context resets.
  (PR #4540)

• Added CartesiaTurnsSTTService for streaming speech-to-text against the Cartesia Streaming ASR v2 (Ink-2) turn-based WebSocket endpoint (/stt/turns/websocket). The server drives turn boundaries via turn.start / turn.update / turn.end messages, which the service translates into UserStartedSpeakingFrame, finalized TranscriptionFrame, and UserStoppedSpeakingFrame. Eager end-of-turn predictions and turn resumes (turn.eager_end and turn.resume) are surfaced via the on_turn_eager_end and on_turn_resume event handlers.
  (PR #4552)

• Added the STTService.supports_ttfs property, which subclasses can override to return False when TTFS doesn't apply to their architecture (e.g. turn-based STTs where the server defines turn boundaries). When False, STTMetadataFrame is broadcast with ttfs_p99_latency=0.0 and the "ttfs_p99_latency not set" warning is suppressed.
  (PR #4585)

Changed

• ⚠️ The development runner now supports all transports (WebRTC, Daily, telephony, plain WebSocket) simultaneously from a single server. The /start endpoint accepts a "transport" field to select the transport per-request; omitting -t at startup enables all transports instead of defaulting to WebRTC. The Daily browser-redirect route moved from GET / to GET /daily.
  (PR #4442)

• Changed the default model for RimeTTSService and RimeHttpTTSService from arcana to coda. Code that relied on the implicit default should set model="arcana" explicitly to preserve previous behavior.
  (PR #4511)

• OpenRouter LLM service now defaults to openai/gpt-4.1.
  (PR #4513)

• OpenRouter LLM requests now convert developer messages to user messages by default for broader model compatibility. Override this by subclassing OpenRouterLLMService or setting llm.supports_developer_role = True for models that support the developer role.
  (PR #4513)

• SonioxSTTService now applies settings updates (e.g. via STTUpdateSettingsFrame) using a graceful reconnect instead of a hard disconnect/reconnect, preserving the service's reconnect retry behavior.
  (PR #4521)

• Updated the default p99 TTFS latency values for Smallest AI, Mistral, and XAI STT so turn stop timing uses measured values instead of the conservative fallback.
  (PR #4522)

• Updated the development runner startup banner to show the prebuilt client URL once and list enabled or disabled transports with install hints.
  (PR #4524)

• Services and transports with missing optional dependencies now raise ImportError instead of a bare Exception when their module is imported without the required extra installed. The original ModuleNotFoundError is preserved as __cause__, so code that wraps these imports can now use except ImportError: cleanly instead of except Exception:.
  (PR #4525)

• Bumped pipecat-ai-prebuilt to 1.0.1 in the runner extra, updating the prebuilt client UI served by the development runner.
  (PR #4531)

• Replaced the transformers.WhisperFeatureExtractor dependency in LocalSmartTurnAnalyzerV3 with a vendored numpy-only implementation, reducing peak RSS at import from ~566 MB to ~60 MB and cold-start time from ~5.0 s to ~0.3 s. Behavior is numerically equivalent (matches the reference numpy code path within 1e-5 absolute tolerance; ONNX model output is bit-identical on representative inputs).

  • Smart Turn v3 no longer imports transformers at module load.
  • Prepares the ground for making transformers an optional dependency in a future release.
  • The vendored STFT is vectorized via numpy.lib.stride_tricks.sliding_window_view + batched np.fft.rfft, cutting _power_spectrogram runtime by ~55% (~4.0 ms → ~1.8 ms per call on a typical 8-second segment at 16 kHz) while preserving the same parity tolerances against the reference implementation.
    (PR #4536)

• ⚠️ Renamed the RTVI UI Worker Protocol's vocabulary from the pipecat-subagents task/agent terms to Pipecat's native job/worker. This spans the wire messages (ui-task → ui-job-group, ui-cancel-task → ui-cancel-job-group), their envelope kinds and fields (task_id → job_id, agents/agent_name → workers/worker_name), the paired Python models/frames (UITask* → UIJobGroup*, RTVIUITask*Frame → RTVIUIJobGroup*Frame), and the @pipecat-ai/client-js / client-react APIs (RTVIEvent.UITask → UIJobGroup, cancelUITask → cancelUIJobGroup, useUITasks → useUIJobGroups, UITasksProvider → UIJobGroupsProvider). These primitives shipped in 1.2.0 but were never documented, so no real consumers are affected.
  (PR #4540)

• transformers is no longer a base dependency, so pip install pipecat-ai no longer pulls it in. This follows Smart Turn v3 dropping its transformers import; the only remaining users (the deprecated LocalSmartTurnAnalyzerV2/CoreML analyzers and the Moondream service) already require the local-smart-turn and moondream extras, which continue to install transformers.
  (PR #4546)

• Widened the deepgram extra to deepgram-sdk>=6.1.1,<8 so installations can resolve to either deepgram-sdk 6.x or 7.x. DeepgramSTTService now handles the agent_rest keyword argument that deepgram-sdk 7.2.0 added to DeepgramClientEnvironment, so custom base_url configuration keeps working on both 6.x and 7.x.
  (PR #4565)

• Dropped the upper bound on the websockets-base extra (websockets>=13.1) so downstream deployments can resolve to websockets 16.x and beyond. Pipecat's websockets usage relies only on the modern websockets.asyncio API plus a handful of public symbols, all of which are retained in 16.x.
  (PR #4565)

• Changed the default voice for GradiumTTSService to _6Aslh2DxfmnRLmP.
  (PR #4569)

• InworldRealtimeLLMService now defaults the STT model to inworld/inworld-stt-1.
  (PR #4573)

Deprecated

• FrameProcessor.pipeline_task is deprecated; read FrameProcessor.pipeline_worker instead. The old name still works but emits a DeprecationWarning and will be removed in a future release.
  (PR #4493)

• Passing a worker to WorkerRunner.run() is deprecated. Register the worker with WorkerRunner.add_workers() before calling run() instead. The worker argument still works but emits a DeprecationWarning and will be removed in a future release.
  (PR #4493)

• PipelineTask, PipelineTaskParams, and the pipecat.pipeline.task module have been renamed to PipelineWorker, WorkerParams, and pipecat.pipeline.worker. The old names still resolve (the module re-exports the new symbols) but constructing PipelineTask / PipelineTaskParams emits a DeprecationWarning; they will be removed in a future release.
  (PR #4493)

• PipelineRunner has been renamed to WorkerRunner and moved to pipecat.workers.runner, since the runner now runs workers (of which PipelineWorker is one kind), not just pipelines. Import WorkerRunner from pipecat.workers.runner. The old pipecat.pipeline.runner module still re-exports both names, and PipelineRunner still works as a subclass alias, but it emits a DeprecationWarning and will be removed in a future release.
  (PR #4589)

Removed

• Removed the unsupported Georgian (Language.KA) language mapping from SonioxSTTService.
  (PR #4521)

Fixed

• Fixed Azure TTS last word being missed by observers and RTVI UI. The completion signal was racing with word timestamp processing, causing the final word's TTSTextFrame to arrive after TTSStoppedFrame. Completion is now routed through the word boundary queue to ensure all words are processed before signaling stream end.
  (PR #4306)

• Fixed skipped TTS frames (e.g. code blocks filtered via skip_aggregator_types) being emitted to the assistant context immediately instead of waiting for preceding spoken frames to finish. They now hold their position in the frame sequence and are flushed only after all earlier spoken sentences are complete, keeping context ordering correct.
  (PR #4380)

• Fixed Cartesia word timestamps leaking SSML tag text (e.g. <spell>, <emotion>, <break>) into word entries. Tags are now stripped before processing, so word-to-text attribution remains accurate when SSML markup is present in the TTS input.
  (PR #4380)

• Fixed BaseOutputTransport reordering frames that share the same presentation timestamp. Frames with equal PTS values are now emitted in insertion order, preventing subtle audio/text sequencing bugs when multiple frames arrive at the same time.
  (PR #4380)

• Fixed TTSTextFrame entries losing their original text structure when word timestamps are enabled. Each TTSTextFrame now carries a raw_text field containing the corresponding span of the original LLM-produced text (including pattern delimiters such as <card>4111 1111 1111 1111</card>), so the assistant context receives properly-tagged content rather than the cleaned words returned by the TTS provider. Also handles words that straddle two sentence boundaries by splitting them and attributing each part to its correct source frame.
  (PR #4380)

• Fixed PipelineTask.cancel() hanging when cancellation is requested before the initial StartFrame reaches the pipeline sink.
  (PR #4455)

• Fixed SmallWebRTCClient.read_audio_frame and read_video_frame busy-looping on MediaStreamError. When a track raises MediaStreamError, the generator now clears the track reference (_audio_input_track / _video_input_track / _screen_video_track) so the loop parks on the existing is None gate instead of re-entering recv() at ~100 Hz on a permanently-failed track. Renegotiation still resumes seamlessly: when _handle_client_connected reassigns a fresh track, the loop picks up frames from the new track.
  (PR #4491)

• Fixed ElevenLabsSTTService crashing when language was passed as None. When language is not set, the service now lets ElevenLabs auto-detect the audio language.
  (PR #4507)

• Fixed NvidiaSTTService so unexpected gRPC stream drops reconnect cleanly using the active audio iterator, while service shutdown and cancellation still close that iterator and stop the streaming worker without leaving it stuck waiting for more audio.
  (PR #4512)

• Fixed websocket STT connection setup failures so services clear stale websocket state and emit non-fatal error frames, allowing ServiceSwitcher failover to keep agents running.
  (PR #4514)

• Fixed ElevenLabsTTSService and ElevenLabsHttpTTSService inserting unwanted spaces between words when synthesizing Chinese or Japanese. Word timestamps for these languages already include their own spacing, so they are now forwarded with includes_inter_frame_spaces=True to avoid double-spacing in transcripts and context.
  (PR #4517)

• Fixed the development runner so missing optional transport dependencies disable only their related routes instead of failing startup in all-transport mode.
  (PR #4524)

• Fixed a race in ElevenLabsTTSService where the periodic keepalive could be sent for a new turn's context before that context's voice_settings initialization message, causing ElevenLabs to close the WebSocket with a 1008 policy violation (voice_settings field must be provided in the first message ...). The keepalive now only targets a context once its context-init has been sent.
  (PR #4527)

• Switched BaseSmartTurn from time.time() to time.monotonic() for its three internal interval-math sites (audio-buffer timestamps, speech-start tracking, and the pre-speech buffer-trim loop). Wall-clock time can step forward or backward when NTP adjusts the system clock, which would silently corrupt the buffer trim (prune everything / prune nothing) and the speech-window extraction. The corrected primitive is monotonic and matches the existing time.perf_counter() usage already in place for inference-latency metrics.
  (PR #4542)

• Fixed SOXRAudioResampler and SOXRStreamAudioResampler ignoring the configured quality setting. Both resamplers were hardcoded to VHQ, which meant RNNoiseFilter's resampler_quality argument (defaulting to QQ for low-latency real-time use) had no effect. The resamplers now honor the configured quality, with VHQ retained as the default.
  (PR #4551)

• Fixed GeminiLiveLLMService (and GeminiVertexLiveLLMService) crashing with 'ContextWindowCompressionParams' object has no attribute 'get' when context_window_compression was supplied through the settings API (e.g. settings=GeminiLiveLLMService.Settings(context_window_compression=ContextWindowCompressionParams(...))). The setting is now handled whether it arrives as a ContextWindowCompressionParams object or as a dict.
  (PR #4563)

• Fixed AudioBufferProcessor concatenating utterances separated by a silent gap. When no user audio arrives for more than 200 ms, silence proportional to the wall-clock gap is now inserted into the user buffer; the same fix is applied symmetrically to the bot buffer, so two bot utterances spoken seconds apart (e.g. progressive hold messages played while a slow function call runs) remain temporally separated in the recorded audio.
  (PR #4567)

• InworldRealtimeLLMService no longer logs WARNINGs for unrecognized realtime server events (e.g. response.output_text.done); they are now logged at DEBUG.
  (PR #4573)

• Fixed a spurious ttfs_p99_latency not set, using default 1.0s warning emitted by turn-based STT services (CartesiaTurnsSTTService, DeepgramFluxSTTService) at pipeline start. These services have no meaningful "speech end → final transcript" interval to measure, because the server defines turn boundaries directly.
  (PR #4585)

Performance

• BaseSmartTurn now stores raw int16 PCM views in its audio buffer and defers the float32 conversion to the once-per-turn segment extraction, eliminating ~50 per-frame numpy allocations per second per analyzer. Output is bit-identical to the previous per-frame conversion path because int16 → float32 / 32768.0 distributes over concatenation; subclasses (LocalSmartTurnAnalyzerV3, LocalCoreMLSmartTurnAnalyzer, HttpSmartTurnAnalyzer) all receive the same float32 audio_array they did before. Also removes a spurious np.frombuffer(audio_int16, dtype=np.int16) re-wrap that was a no-op view-of-a-view of already-int16 data.
  (PR #4542)

• Reduced the soxr resampling quality preset in LocalSmartTurnAnalyzerV3 from VHQ (~26-tap polyphase) to HQ (~16-tap), cutting resample CPU time by 30–50% on non-16 kHz audio sources (~3–10 ms saved per turn at 24/48 kHz). Pipelines already delivering 16 kHz audio are unaffected — the existing actual_rate == _MODEL_SAMPLE_RATE fast path skips resampling entirely. The two quality presets differ in filter length, not cutoff or interpolation semantics; on a Whisper-style log-mel feature representation the audible difference sits well below the mel filterbank's quantization noise floor, so model predictions are unchanged on representative inputs.
  (PR #4542)

Changed

• Changed the default WebSocket endpoints for GradiumSTTService and GradiumTTSService to the region-neutral wss://api.gradium.ai/api/speech/asr and wss://api.gradium.ai/api/speech/tts. Gradium now automatically routes traffic to the nearest endpoint. Override the url to pin to a specific region.
  (PR #4500)

Fixed

• Fixed bot hangs when filter_incomplete_user_turns was enabled and the LLM responded by calling a tool. The user turn never finalized, so the assistant aggregator gated the tool-result context push and the LLM continuation never ran. Tool calls now finalize the turn the moment they start, before the function dispatches.
  (PR #4501)