PyPI · #3911 most downloaded on PyPI

pipecat-ai

1.12.01.1M downloads/mopipecat-ai/pipecat ↗New releases in your daily digest, changelog included.

An open source framework for voice (and multimodal) assistants

Last release 8 days ago
26 Sep 2026
Ships on a steady schedule
a new release about every 2 weeks
Nearly every release is documented
notes for 60 of the last 60 stable releases
2 versions withdrawn
withdrawn after publishing
2 years old
115 releases · first in 2024

Release timeline

115 releases since 2024

202420252026

Releases

1.12.026 Sept 2026Deprecated3 sources agree
Deprecated UninterruptibleFrame . A frame class that should be uninterruptible by default declares interruptible: bool = field(default=False, init=Fal…
Added

TTS sentence aggregation follows Settings.language, defaulting to English when unspecified. Language updates apply with TTS settings.
(PR #5737)

Added TTSService.pronunciation_transform_ipa(), which builds a text transform from a word-to-IPA mapping so a voice says names and terms it would otherwise guess from spelling. Each matched word is replaced with the service's format_pronunciation() output, so one mapping works with any service that supports pronunciation hints; a word the service cannot use is reported once when the transform is built and spoken as written. Cartesia writes IPA as inline phoneme blocks, ElevenLabs as SSML <phoneme> tags (read by eleven_flash_v2 and eleven_turbo_v2, and on the WebSocket service only with enable_ssml_parsing=True; on any other setup the transform is skipped with a warning and words are spoken as written), Eleven v3 and Inworld as IPA between slashes, and Deepgram Aura-2 as an inline pronunciation object. Deepgram Flux has no pronunciation markup, so its words are spoken as written. pipecat.utils.text.phonemes holds the IPA parsing and normalization the formatters share. Register the transform last in text_transforms, so no later transform rewrites the markup it inserts.

pronounce = CartesiaTTSService.pronunciation_transform_ipa({"Metformin": "mɛtˈfɔɹmɪn"}) tts = CartesiaTTSService( text_transforms=[ ("*", strip_markdown), ("*", pronounce), # last, so nothing rewrites its markup ], )

(PR #5798)

Added client-mode reconnect to MOQTransport: when the relay session drops, the transport redials with backoff for as long as MOQParams.connection_timeout allows the client to be missing, keeps its broadcast across dials, and reports the client gone only when that time is up. A relay that vanishes without closing the session is detected by watching the connection's traffic counters, rather than waiting out the QUIC idle timeout. Each side now sends a session-ending marker on its transcript before leaving; a peer whose tracks end without it is treated as an outage rather than a hangup. on_disconnected and on_connected now fire for each relay session, so a redial shows up as a disconnect followed by a connect; end the call from on_client_disconnected, which fires only once the client is gone.
(PR #5800)

Added MOQRunnerArguments.relay_url for dialing a relay by its full URL, query string included. host and port are optional when it is set, and create_transport passes it through to MOQParams.relay_url. When host or port is given as well, relay_url wins and a warning is logged.
(PR #5800)

Added interruptible to every frame: True by default and False by default for UninterruptibleFrame subclasses, and what the frame queue, the processor's interruption handling and the speculation gate decide by. Set it on a frame before pushing it to keep that one frame through an interruption, or to let one frame of a protected type be dropped, without a frame class of your own.
(PR #5835)

JudgeVerdict now has a confidence, from 0 to 1, saying how sure the judge is. With Jev the number is calibrated. With an LLM it is the LLM's own guess.
(PR #5857)

The eval judge can now use a classifier. Put a factory: in the scenario's judge.eval: block that returns a BaseClassifier, or an LLM service as before. Our release evals judge with Jev this way, through evals/judges.py. Jev answers each check in a few hundred milliseconds and says how sure it is. It needs the jev extra and TYPESAFE_API_KEY.
(PR #5857)

Added allow_continue to a scenario's judge.eval: block and to EvalJudge. Set it to false when every reply you judge is a final answer. The judge then answers yes or no and never continue, so it never waits for more text that isn't coming.
(PR #5857)

Added LLMClassifier, which answers classifier questions with any Pipecat LLM service that supports run_inference(). All the questions about one state go to the LLM in one call, and the LLM replies with one JSON object holding an answer per question. It waits timeout seconds for the LLM, 10 by default, and raises ClassifierError if the LLM does not answer in time or the call fails.

classifier = LLMClassifier(llm=OpenAILLMService(model="gpt-4o-mini")) results = await classifier.yes_no( "Hi, you've reached Dana. Leave a message.", {"voicemail": YesNoQuestion(instructions="is this a voicemail greeting?")}, ) results["voicemail"].is_yes # True results["voicemail"].probability # 0.97

(PR #5863)

Classifiers report metrics through an on_metrics event: after every call, the time it took as ProcessingMetricsData and, for JevClassifier, the tokens it used as LLMUsageMetricsData. A classifier cannot push frames, so its owner puts the data in a MetricsFrame.
(PR #5863)

Added Pipecat Classifiers. A classifier is a small object that answers typed questions about some state. A question is a YesNoQuestion, a ChoiceQuestion among options, or a ScoreQuestion on a scale. Questions are asked by name, several about one state at once. ask() takes any mix of kinds, and yes_no(), choice() and score() take questions of one kind and return typed results: a YesNoResult with the probability of yes and is_yes, a ChoiceResult with the choice and a probability per option, and a ScoreResult with the score on the scale and a probability per level. JevClassifier answers them through Jev, TypeSafe's classification model, in one request. A choice question can have at most 255 options, Jev's limit (JEV_MAX_CHOICE_OPTIONS). Several classifiers can share one JevClient, which handles the HTTP/2 connection, auth, retries and token accounting. Install with uv add "pipecat-ai[jev]". examples/features/features-classifiers.py asks one question of each kind.

classifier = JevClassifier(api_key=os.getenv("TYPESAFE_API_KEY")) results = await classifier.choice( "I'd like to book a table for, um", { "turn": ChoiceQuestion( instructions="is the user's turn over?", options={ "complete": "the user finished", "short": "a brief pause", "long": "asked for time", }, ) }, ) results["turn"].choice # "short"

(PR #5863)

Added a read-only settings property to AIService, so code outside a service can read its current settings, such as the model. Settings are still changed through a ServiceUpdateSettingsFrame.
(PR #5863)

Added an optional response_schema argument to LLMService.run_inference(), a JSON schema the reply must follow. OpenAI, Anthropic and Google have the provider enforce it, so the reply is JSON text in that shape. A service or model that cannot enforce one, such as Bedrock, DeepSeek or an OpenAI model before gpt-4o-mini, ignores it with a warning, and supports_response_schema says whether a service can at all.

reply = await llm.run_inference(context, response_schema=schema)

(PR #5876)

Added LLMUserAggregatorParams.empty_user_turn, an EmptyUserTurnConfig for user turns that end with no transcript, such as a cough, background noise, or speech the STT could not recognize. It is on by default, even when empty_user_turn isn't passed, and treats two cases differently:

Interrupted: the turn cut the bot off. The bot would otherwise stay silent mid-response, so the aggregator appends a developer message (interrupted_prompt) and runs the LLM once, and the bot asks the user to repeat or picks up where it left off. On by default.

Idle: the bot had finished and was waiting for the user. The conversation isn't stuck, and answering what may be noise would be intrusive, so these turns get no answer unless idle_prompt is set.

LLMUserAggregatorParams( empty_user_turn=EmptyUserTurnConfig( idle_prompt=( "The user may have said something, but it was not " "recognized. Briefly ask them to repeat it." ), ), )

Setting interrupted_prompt=None turns off the interrupted case, and empty_user_turn=None turns off both. max_consecutive_recoveries (default 1) limits how many empty turns in a row get an answer.
(PR #5907)

UIWorker can now use a classifier for small decisions about the screen, with no LLM turn: which element the user means (which_element), whether something is true on the screen (check_screen), which elements match a description (select_elements), whether a UI event deserves a reply (should_respond), and act on an element named in words. Pass a classifier, for example a JevClassifier. If you don't, the worker's own LLM answers through an LLMClassifier, which works but is slower.

worker = MyUIWorker("ui", llm=OpenAILLMService(api_key=...), classifier=JevClassifier(api_key=...)) ref = await worker.which_element("the checkout button")

(PR #5909)

A voice LLM can now ask a UIWorker about the screen without ever seeing it. The worker has a built-in screen job with an action: find (which element a description means), check (whether something is true on the screen), select (which elements match a description), list (what is on the screen), selection (the text the user has selected) and click, scroll_to, highlight, select_text or fill (do that to the element a description means). Every answer is short data, never the page. screen_tools("ui") gives the voice LLM the tool that sends it.

context = LLMContext(tools=screen_tools("ui"))

(PR #5909)

UIWorker now has a snapshot property with the latest page snapshot and a selection property with the text the user has selected, so a subclass can read the screen with plain code.
(PR #5909)

Added RTVIFunctionCallReportLevel.ARGUMENTS, between NAME and FULL. It reports the function name and the arguments, and no result.
(PR #5918)

Changed

InworldRealtimeLLMService now sends providerData.auto_tool_response=false, leaving Pipecat responsible for requesting continuation after submitting a tool result.

InworldRealtimeLLMService now generates a unique session key for every connection, including connections opened within the same millisecond.
(PR #5116)

Sentence aggregation uses self-contained sentencex rules without NLTK data downloads or tokenizer warm-up. Language-specific sentence boundaries may differ from Punkt.
(PR #5737)

MOQTransport numbers every transcript record with seq and epoch and drops replayed records on subscribe, so a reconnect on either side no longer redelivers the whole RTVI log and re-fires client-ready. Both fields are stripped before the message reaches the pipeline; records without them pass through unchanged, so older peers keep working.
(PR #5800)

MOQTransport failures now reach the pipeline as an ErrorFrame with an error category, in addition to the on_error event. A relay that refuses the token, at the dial or by closing the session as unauthorized, is not retried and marks the transport unusable, as does a relay that cannot be reached within connection_timeout.
(PR #5800)

The moq extra requires moq-rs 0.4.6 or later. moq-rs is the client library MOQTransport dials a relay with; moq-relay is the separate server. The transport is tested against moq-relay 0.14. A relay from an incompatible release line can accept the connection and still drop the transcript track without an error.
(PR #5800)

MOQParams.connection_timeout is now the one limit on how long the client may be missing, and defaults to 60 s (was 30 s). It bounds the wait for the client to join and, in client mode, an outage: the time from the relay session dropping until the client's data flows again, across every redial. Set it longer than the time your load balancer takes to fail a dead relay out.
(PR #5800)

XAISTTService now sends its model setting to xAI, and defaults to grok-voice-transcribe-2.0. Previously no model was sent, so xAI used its server default, grok-voice-transcribe-1.0. To keep the previous model, pass settings=XAISTTService.Settings(model="grok-voice-transcribe-1.0").
(PR #5847)

The local-smart-turn and moondream extras now require transformers>=5.10.0. The moondream extra's accelerate, einops, pyvips and timm pins are relaxed to a lower bound with a major-version cap.
(PR #5850)

EvalJudge now decides every verdict with a classifier. By default that is an LLMClassifier over the LLM the judge.eval: block names, so existing scenarios work as before. A classifier gives no reasons, so the judge asks an LLM, the explainer, for the reason behind every no and every verdict below explain_below. The explainer is the judging LLM unless an explainer: block names another one, and explainer: false turns reasons off. The classifier's verdict always stands.
(PR #5857)

pipecat eval suite now keeps concurrency runs going at all times, taking the next run in manifest order. Before, each entry ran its scenarios one at a time on one slot, so a manifest with fewer entries than slots left slots idle, and one entry with ten scenarios ran them one by one. An entry whose provider rate-limits sets its own concurrency: to cap its runs in flight, as the turn-completion manifest does.
(PR #5857)

The mcp extra now requires mcp 2 (mcp[cli]>=2.1.1,<3). MCPClient works the same; it no longer runs on the 1.x SDK.
(PR #5857)

An LLM judge now classifies instead of answering a prose prompt: it gets the conversation as structured state and picks one of the outcomes. A borderline reply can get a different verdict than before. A simulation is judged one bot turn per call instead of the whole run in one call, so it costs more calls.
(PR #5857)

VoicemailDetector now takes a classifier and is a single processor instead of a parallel pipeline with its own LLM. It asks the classifier after each transcription and acts once the caller has been quiet for decision_timeout (default 1 s), when the latest answer decides. No verdict acts on a fragment: "hi, this is Sam" is what a person says and how a greeting starts, and only the silence that follows tells them apart. Then the held-back speech is released or dropped as before. The on_voicemail_detected and on_conversation_detected handlers receive the detector itself.

# Before detector = VoicemailDetector(llm=OpenAILLMService(api_key=...)) # After detector = VoicemailDetector(classifier=JevClassifier(api_key=...))

(PR #5869)

Observers created with observe_every_push=False are told about a frame once, on its first push, instead of on every push by every processor that passes it along. The built-in observers that handle a frame once do so, and FramePushed.first_push tells the first push from the ones that follow for the ones that observe every push.
(PR #5908)

A UIWorker now answers a respond job with the reply its LLM writes, so the smallest UI worker is a UIWorker with an LLM and a system prompt. A @tool that calls respond_to_job still answers instead when it needs to, for example to speak the answer through TTS.
(PR #5909)

UI_SNAPSHOT_EVENT_NAME and UI_CANCEL_JOB_GROUP_EVENT_NAME are now public in pipecat.bus.ui. They are the bus event names of the client's screen snapshot and of the client asking to cancel a job group.
(PR #5909)

Deprecated

Deprecated UninterruptibleFrame. A frame class that should be uninterruptible by default declares interruptible: bool = field(default=False, init=False) instead. The marker still sets the flag until it is removed in 2.0.0.
(PR #5835)

Passing an LLM service to EvalJudge, as the first argument or as service=, is deprecated and will be removed in 2.0.0. Pass EvalJudge(LLMClassifier(llm=service), explainer=service) instead, which is what the old form did.
(PR #5857)

Deprecated the startup warming timings reported by StartupTimingObserver, all removed in 2.0.0: the StartupTimingReport.warmup field, StartupWarmupTiming, StartupWarmup, and the on_startup_warmup() hook on BaseObserver, WorkerObserver and StartupTimingObserver. Pipecat warms no deferred imports while the pipeline sets up, so warmup is always None and the hook is never called. The rest of the report — the phase totals and the per-processor timings — is unchanged.
(PR #5859)

VoicemailDetector's llm and custom_system_prompt parameters are deprecated. Pass a classifier instead. Until removal, the llm is wrapped in an LLMClassifier, with the custom prompt in front of the classifier's instructions.
(PR #5869)

The max_frames parameters of TurnTrackingObserver and UserBotLatencyObserver are now deprecated, removed in 2.0.0. Observers no longer keep a window of the frames they have seen.
(PR #5908)

tts_speak on UIWorker.respond_to_job is deprecated and will be removed in 2.0.0. A UI worker should not speak; respond with the answer and let the voice LLM say it.
(PR #5909)

BaseUIWorker is deprecated and will be removed in 2.0.0. Use UIWorker instead. UIWorker now reports its job groups to the client on its own, so instead of a separate BaseUIWorker dispatcher, dispatch job groups from a @job handler in your UIWorker.
(PR #5909)

ReplyToolMixin is deprecated and will be removed in 2.0.0. Use screen_tools instead: the voice LLM asks the UI worker through the screen job and says the answer itself.
(PR #5909)

Removed

Removed NotifierGate, ClassifierGate, ConversationGate and ClassificationProcessor from pipecat.extensions.voicemail.voicemail_detector, and the CLASSIFIER_RESPONSE_INSTRUCTION and DEFAULT_SYSTEM_PROMPT attributes of VoicemailDetector. They were parts of the old parallel-pipeline detector and its prompt.
(PR #5869)

Fixed

Fixed an LLMService regression where re-advertising a tool with the same name but a different handler (for example, a per-node handler in Pipecat Flows) silently kept the previous handler bound. Auto-registered handlers are now rebound when the advertised handler changes, while explicit register_function registrations are still left untouched.
(PR #4823)

Fixed InworldRealtimeLLMService resending a server-VAD user transcript after a fast tool result context update, which duplicated user turns and responses.
(PR #5116)

InworldRealtimeLLMService now serializes Inworld extensions under providerData instead of provider_data, allowing the server to apply them.
(PR #5116)

Applied NVIDIA STT settings changes to the running stream. NvidiaSTTService rebuilt its recognition config on a settings update but never reconnected, so the open gRPC stream kept transcribing with the previous settings while self._settings reported the new ones.
(PR #5632)

Fixed Flows node transitions where an interruption could leave the LLM running with the previous node's context and tools. The LLMMessagesAppendFrame or LLMMessagesUpdateFrame and the LLMSetToolsFrame a transition queues are now uninterruptible, so they are still delivered.
(PR #5837)

TelnyxFrameSerializer now sends OutputTransportMessageFrame and OutputTransportMessageUrgentFrame to the client as JSON, matching the other telephony serializers. Previously these messages (including RTVI messages) were silently dropped.
(PR #5841)

Fixed the eval harness treating the bot's earlier speech as its reply when a scenario interrupts the bot. Speech from before the interruption is now dropped.
(PR #5846)

Fixed MoondreamService with transformers 5: loading the model on a GPU or Apple Silicon failed with AttributeError: 'HfMoondream' object has no attribute 'all_tied_weights_keys', and a model that did load produced garbage descriptions.
(PR #5856)

Fixed a bridged PipelineWorker reporting every frame from its LLM twice over RTVI, once in its own pipeline and once when the frame crossed the bridge, which doubled every word of a bridged worker's reply in the client's LLM text and in the evals. enable_rtvi now defaults to off for a bridged pipeline worker, which has no client of its own; LLMWorker already did this. Pass enable_rtvi=True to keep it.
(PR #5857)

Fixed TTS word tracking falling out of step on markdown-heavy replies. Most word-timestamp events were dropped with "Dropping word ... not recognised by any slot" warnings, and the rest of the reply arrived in large chunks, so word highlighting stalled and then jumped ahead. Two token shapes triggered it: the period Cartesia adds to the last token of a line (images:., ---.), and a symbol a provider reports differently from the text (ElevenLabs reports → as -) followed by more symbols such as ### or **.
(PR #5866)

Fixed WorkerRunner skipping a worker that another worker added while the runner was still starting, for example a child worker added by a processor during its setup.
(PR #5872)

Fixed pcm_to_wav() dropping complete samples when given a typed memoryview.
(PR #5879)

Fixed is_silence() misclassifying full-scale negative int16 PCM samples as silence.
(PR #5881)

Fixed MOQTransport ending the call when a relay refused a subscription to the peer's broadcast with dropped, which a surviving relay in a mesh does while it still holds a route through a relay that went away. The refusal is now treated as the peer's tracks ending, so the transport retries and redials as it does for any other relay loss.
(PR #5892)

Fixed RTVIObserver remembering the IDs of frames it does not handle, such as every audio frame. Its memory still grows with the frames it handles, but much more slowly.
(PR #5906)

Fixed on_user_turn_idle never firing after a user turn that ended with no transcript, since that turn cancelled the idle timer and nothing restarted it.
(PR #5907)

Fixed observers keeping the IDs of every frame they had seen for the life of the session, the pipeline worker's idle detection included.
(PR #5908)

The runner extra now requires pipecat-ai-prebuilt>=1.2.2. The conversation panel in the prebuilt client UI served by the development runner now keeps autoscrolling while long bot replies stream in.
(PR #5917)

Performance

RTVIObserver now skips audio frames before checking anything else when audio levels are not reported, instead of testing every condition on every push.
(PR #5906)

1.11.018 Sept 2026Deprecated3 sources agree

Deprecated EvalSimulationDriver.transcript() . No replacement: the driver feeds the judge as the conversation happens. Will be removed in 2.0.0. (PR #…

Added

Added BaseAudioResampler.flush() and BaseAudioResampler.reset(), so callers can mark stream boundaries themselves rather than relying on SOXRStreamAudioResampler's inactivity timeout. flush() returns the audio the resampler is still holding, for a stream that has ended; reset() discards it, for one that was abandoned. Both default to no-ops for resamplers that keep no state between calls.
(PR #5386)
Scripted eval expectations on function_call accept an eval:: each matched call is judged by name and arguments, under a judge prompt of its own, so a scenario can check what args: cannot match verbatim ("a session about OpenTelemetry tracing, submitted for Jennifer Smith"). A rejected call fails the turn with kind judge_no. EvalJudge gains evaluate_call().
(PR #5713)
AWSTranscribeSTTService.Settings gained partial_results_stability, which selects AWS Transcribe's "high", "medium" or "low" interim-result stability. It still defaults to "high".
(PR #5743)
Added an effect setting to AzureTTSService and AzureHttpTTSService, which passes Azure's <voice effect> audio effect processor (eq_car, eq_telecomhp8k) through to the synthesis request.
(PR #5744)
Added a reasoning_effort field to GroqLLMService.Settings, which passes Groq's reasoning effort control through to the completion request. Accepted values vary by model: "low", "medium" and "high" on the GPT-OSS models and qwen/qwen3.8-27b, "none" and "default" on the Qwen models. Setting "none" also keeps a Qwen model's <think> reasoning out of the spoken response.
(PR #5748)
Added a delay field to OpenAIRealtimeSTTService.Settings, OpenAI's latency-versus-accuracy control for how long the model waits before emitting transcription text (minimal, low, medium, high, xhigh). Supported by gpt-realtime-whisper.
(PR #5750)
Added a keywords field to OpenAISTTService.Settings, which passes product names, acronyms and other specialized terms to OpenAI's transcription request as hints. Supported by gpt-transcribe.
(PR #5750)
Added math_notation to SmallestTTSService.Settings, Smallest's opt-in flag for reading digit-flanked math operators as words (2 + 2 as "two plus two").
(PR #5751)
Added an initial_prompt field to WhisperSTTService.Settings, which gives Faster Whisper preceding context for each segment, steering the transcript's style, punctuation and spelling.
(PR #5755)
SarvamRealtimeSTTService now accepts Sarvam's second-generation realtime model, saaras:v4, via settings=SarvamRealtimeSTTService.Settings(model="saaras:v4"). The default remains saaras:v3-realtime.
(PR #5758)
Added video output to LiveKitTransport. With video_out_enabled, the transport publishes a camera track from RGB, RGBA, BGRA or ARGB frames. video_out_codec selects the codec, and the new LiveKitParams.video_out_max_bitrate caps the bitrate together with video_out_framerate.
(PR #5770)
Added a bot-llm-marker RTVI message that RTVIObserver sends when its new bot_llm_marker_enabled parameter is on (off by default), reporting the turn-completion marker the bot's LLM produced and what it meant: complete, short or long.
(PR #5777)
Scripted eval scenarios can assert on the turn-completion marker the bot's LLM produced with a new llm_marker event and a marker: field naming its meaning: complete, short, long, or incomplete for either of the last two. The bot reports its markers only when a scenario asks for them, so clients never see them by default.
```
- user: "Let me think about it, hmmm"
  expect:
    - event: llm_marker
      marker: incomplete
```
(PR #5777)
A scenario file lists its scenarios under scenarios:, so one file can hold many short cases of the same behavior. Any key a scenario can have may sit at the top of the file as the default for all of them, and a scenario that sets the same key replaces it whole. turns: at the top with one entry per judge or modality runs the same conversation under each; persona: at the top with a goal: per entry sends the same caller on different errands. Scripted and simulated scenarios can share a file. Each scenario runs on its own, against its own bot, and is named <file>/<scenario>; the suite's -s filter accepts that name or either half of it. EvalScenarioFile.load() reads a file and holds every scenario in it.
```
name: turn_completion
judge: !include ../judge_text.yaml

scenarios:
  - name: short_answer
    turns:
      - user: "Japan."
        expect:
          - event: response
  - name: cutoff
    turns:
      - user: "I'd go to Japan because"
        expect:
          - event: response
            absent: true
            within_ms: 3000
```
(PR #5780)
GeminiLiveLLMService supports gemini-3.8-live and gemini-3.8-live-extended-thinking, including NON_BLOCKING function calling (the family's default). Synchronous tools are declared BLOCKING on gemini-3.8-live; gemini-3.8-live-extended-thinking accepts only NON_BLOCKING and logs a warning for them. The thinking model requires a thinking_level and gets LOW when none is configured.
(PR #5781)
GeminiLiveLLMService holds the bot turn open while Gemini reports interaction_status: IN_PROGRESS, so a reply that spans several turn_complete messages is recorded as a single assistant turn. Requires google-genai>=2.19.0 (now the google extra's floor); older SDKs log a warning.
(PR #5781)
Added name: to a manifest entry, its label in the display, the -p filter, results.jsonl (a new name field beside bot) and the artifact file names, which carry the bot path and then the name. It defaults to the bot: path, for entries that share a bot and differ only in their runner_body:. Two entries may no longer run the same scenario under one label.
(PR #5786)
Added concurrency: to a manifest entry, capping how many runs of that bot execute at once under the suite's own concurrency, for providers that rate-limit concurrent connections.
(PR #5786)
Added text_excludes: to a scripted scenario's expectations, the mirror of text_contains: the event's text must not hold the given substring, for a marker the LLM let slip into its reply. It fails with kind text_present.
(PR #5786)
Added data: to a manifest entry's runner_body:, the bot's runner-args body written inline as a mapping; the suite writes it to a file and passes it as --runner-body. The runner reads a body file as YAML, so it may be YAML or JSON.
(PR #5786)
Added EvalScriptTurnResult.expectations, one EvalExpectationResult per expectation the turn resolved, with whether it passed and what it matched (the marker of an llm_marker, a function call's signature, a reply's text). The suite's results.jsonl records carry it too.
(PR #5786)
Added LLMMarkerResponseFrame, pushed by UserTurnCompletionLLMServiceMixin when a response ends with the raw text the LLM produced, the marker it read and the markers it recognizes. The RTVI observer sends it as bot-llm-marker when bot_llm_marker_enabled is on, so the eval harness's llm_marker event now carries raw and markers, and a scripted scenario can check the marker protocol with marker_first, markers and text_after on an llm_marker expectation.
(PR #5786)
AssemblyAISTTService's language_codes steering now accepts Urdu, Russian, Korean, Catalan, Galician, Romanian, Estonian, Persian, Cantonese, Afrikaans, Marathi, Zulu, Xhosa, and Norwegian Nynorsk, matching the full set universal-3-5-pro and universal-3-6-pro support.
(PR #5808)
Added speed to the eval harness's speech: block, Kokoro's rate multiplier, so a scenario can make the synthesized user talk faster or slower. Cached user audio is keyed by the speed too.
(PR #5813)

Changed

BasetenLLMService defaults to zai-org/GLM-5.3. Set settings=BasetenLLMService.Settings(model=...) to use a different model.
(PR #5730)
DeepgramFluxSTTService now documents its default url as Deepgram's Flux endpoint, wss://api.deepgram.com/v2/listen.
(PR #5739)
DeepSeekLLMService defaults to deepseek-flash, DeepSeek's current name for V4.1 Flash. The previous default, deepseek-v4-flash, names a retired model and is only temporarily routed to V4.1 Flash. Set settings=DeepSeekLLMService.Settings(model=...) to use a different model.
(PR #5746)
Added a mode field to the OpenAI Responses ReasoningConfig, selecting the standard or pro reasoning mode on models that offer one, such as gpt-5.6.
(PR #5749)
SonioxTTSService settings now include client_reference_id, the identifier Soniox records with each request in its usage logs, matching the field SonioxSTTService already exposes.
(PR #5752)
Added max_session_duration to LiveAvatarNewSessionRequest, capping a HeyGen LiveAvatar session at a given number of seconds.
(PR #5753)
InworldRealtimeLLMService docstrings now use openai/gpt-4.1-mini in their examples instead of openai/gpt-4.1-nano, which OpenAI shuts down on 2026-10-23.
(PR #5754)
RimeHttpTTSService now sends noTextNormalization to Rime, so the setting takes effect on Mist requests.
(PR #5756)
SarvamLLMService accepts deepseekv4-flash, the DeepSeek V4 Flash model Sarvam serves on /v2 with a 1M-token context window, tool calling, and reasoning.
(PR #5757)
pipecat init React clients (Vite and Next.js) now use Pipecat UI instead of @pipecat-ai/voice-ui-kit. The generated client renders the Pipecat UI console: connect flow, transcript, metrics, device and session info, and a live event stream. Components are installed as source under src/components/pipecat, and components.json points at the @pipecat registry, so npx shadcn@latest add @pipecat/<component> adds more. The vanilla JavaScript client is restyled to match.
(PR #5767)
pipecat eval suite now takes its queue round-robin across the manifest's entries within each attempt, so a slow or rate-limited provider holds only its share of the concurrency instead of every slot until its scenarios are done.
(PR #5768)
Rewrote the default instructions that filter_incomplete_user_turns appends to the system prompt, so weaker models follow the ●/◐/○ marker protocol more reliably. Short answers such as "yes" or "Tuesday" are now called out as complete, a mid-sentence fragment is explicitly not a request for help, a continuation is judged together with the fragment before it, and the examples are prose rather than arrows or a transcript, since small models reproduce whatever shape they are shown. Bots that set UserTurnCompletionConfig.instructions are unaffected.
(PR #5768)
SpeechmaticsSTTService now defaults turn_detection_mode to TurnDetectionMode.EXTERNAL, so Pipecat's own VAD drives turn boundaries through finalize(). Set settings=SpeechmaticsSTTService.Settings(turn_detection_mode=TurnDetectionMode.VAD) to keep letting the Speechmatics service run its own VAD and close turns itself.
(PR #5773)
The runner extra now requires pipecat-ai-prebuilt>=1.1.1. The prebuilt client UI served by the development runner moves to @pipecat-ai/client-js 1.13.1, @pipecat-ai/daily-transport 1.6.9, @pipecat-ai/small-webrtc-transport 1.10.8, and @pipecat-ai/websocket-transport 1.7.2.
(PR #5776)
The runner extra now requires pipecat-ai-prebuilt>=1.2.1. The prebuilt client UI served by the development runner is built with Pipecat UI instead of @pipecat-ai/voice-ui-kit, shows the bot video pane in the console, and follows the system light/dark theme on load.
(PR #5799)
The sagemaker and aws-nova-sonic extras now require aws_sdk_sagemaker_runtime_http2>=0.6.0,<0.10 and aws_sdk_bedrock_runtime>=0.6.0,<0.10. The two SDKs share smithy-core, so they move together.
(PR #5807)
Async function calls (cancel_on_interruption=False) now keep settling as ordinary tool results when only assistant output lands in the context while they run, such as filler a tool handler speaks with TTSSpeakFrame. Previously that filler turned the result into a deferred-result message, while the same filler spoken from on_function_calls_started did not. The result is still deferred when a user or developer message arrives first, or when the call sent an intermediate update.
(PR #5820)
resolve_language(..., use_base_code=True) now resolves a regional variant missing from the map to the map's code for its base language when the map has one (en-US becomes "eng" for a map with Language.EN: "eng"), falling back to the base code ("en") otherwise. This changes the code sent for unmapped regional variants by Meta STT, Pocket TTS, Rime TTS, Camb TTS, Kokoro TTS, XTTS, Neuphonic TTS, and xAI TTS; for example, Rime sends "ara" for Language.AR_AE instead of "ar".
(PR #5824)
pipecat eval suite runs each manifest entry's scenarios back to back on one concurrency slot, so an entry's next scenario starts as soon as its previous one ends and a slow provider holds no more than one slot. A repeated sweep is attempt-major, every entry's first attempt before any entry's second. An entry's concurrency: now means how many slots it may hold at once, one by default.
(PR #5825)

Deprecated

Deprecated EvalSimulationDriver.transcript(). No replacement: the driver feeds the judge as the conversation happens. Will be removed in 2.0.0.
(PR #5765)
Deprecated the transcript parameter of EvalJudge.evaluate_run(). The judge keeps the conversation it judges: feed it with add_user_message, add_assistant_message and the new add_tool_call, and call evaluate_run(criteria, success). Will be removed in 2.0.0.
(PR #5765)
EvalScriptScenario.load() and EvalSimulationScenario.load() are deprecated and will be removed in 2.0.0. Use EvalScenarioFile.load(), which reads a scenario file of either kind.
(PR #5780)
load_scenario_file() is deprecated and will be removed in 2.0.0. Use EvalScenarioFile.load(), which reads a file and holds every scenario in it.
(PR #5780)
A scenario file whose top level holds turns: or persona: instead of a scenarios: list is deprecated and will stop loading in 2.0.0. It still loads as one scenario under the file's name, with a DeprecationWarning. Wrap the scenario in a scenarios: list with its own name:.
(PR #5780)
Deprecated a manifest entry's bare runner_body: <file>. Use runner_body: {path: <file>} instead. Will be removed in 2.0.0.
(PR #5786)

Fixed

Fixed the output transport dropping speech whenever TTS pauses for more than 200 ms between chunks. This only affected pipelines where the TTS render rate differs from the transport's output rate.
(PR #5386)
Fixed a word-timestamp event that belongs to no sentence being discarded with nothing logged, in streaming mode (a TTS service using TextAggregationMode.TOKEN). This comes up when a provider reports a punctuation mark as its own event after the preceding word already took it. Such an event used to sit in a buffer until the next turn began and was thrown away there; it is now logged and dropped when its own turn ends.
(PR #5681)
Fixed AggregatedFrameSequencer.force_complete emitting the text it force-completes as a TTSTextFrame without a matching AggregatedTextProgressFrame, so the progress view stopped where the TTS provider stopped reporting words while the word frames carried the rest of the turn.
(PR #5681)
Fixed audio eval sessions remaining silent after a text-mode session on the same bot by resetting skip_tts for each connection.
(PR #5722)
A failed MCPClient tool call now gives the LLM the error the call raised, cut to 200 characters. The fixed line Sorry, could not call the mcp tool remains only for a call that raised nothing and returned no text content. Code that matched on the old text must read the new one.
(PR #5734)
CartesiaSTTService now sends keyterm for Cartesia's ink-preview models, which support keyterms alongside ink-2.
(PR #5738)
Corrected the documented range of the Deepgram Flux TTS speed setting: Flux accepts 0.5 to 1.5 in steps of 0.05, not 0.85 to 1.15.
(PR #5740)
Fixed GoogleLLMService failing every request with gemini-3.8-flash. That model rejects the minimal thinking level Pipecat applies by default to Gemini 3 Flash models, so it now gets low, the lowest level it accepts.
(PR #5741)
The Mem0 RAG example's local-configuration snippet now names claude-sonnet-4-6 instead of the retired claude-3-5-sonnet-20240620.
(PR #5742)
Corrected MoonshineSTTService's licensing note: Moonshine's models are MIT in every language and size, except the legacy non-streaming (tiny, base) non-English models, which stay under the non-commercial Moonshine Community License.
(PR #5745)
Updated the Fireworks update-settings example to use nemotron-3-ultra-nvfp4; Fireworks retired gpt-oss-20b from serverless.
(PR #5747)
The OpenAI Responses services no longer send reasoning.effort="none" for gpt-6-astra, which rejects that value. The model now works out of the box.
(PR #5749)
Fixed HeyGenVideoService avatar state logging, which watched for the LiveAvatar LITE event agent.state. LiveAvatar renamed it to agent.state_updated.
(PR #5753)
Fixed the voice-sarvam example's optional voice-switch line, which named anushka, a bulbul:v2 speaker that bulbul:v3 rejects. It now names anand, a bulbul:v3 speaker.
(PR #5759)
Fixed per-tool timeout_secs (and global function_call_timeout_secs) being disarmed by an intermediate FunctionCallResultProperties(is_final=False) update. An async tool that reported progress no longer cancels its own deadline; only a final result clears it.
(PR #5760)
npm run lint in a generated React client failed before linting any file, because the ESLint config used a legacy plugin shape that ESLint 10 rejects. The Vite template now uses the flat react-hooks config, and the Next.js template pins ESLint 9 since eslint-config-next 16 does not run on 10.
(PR #5767)
Fixed SambaNovaLLMService ignoring filter_incomplete_user_turns. Its text bypassed the turn-completion mixin, so incomplete turns were never held and the marker character was spoken by TTS.
(PR #5768)
Fixed LiveKitTransport reporting a disconnect twice when it disconnected, and reporting one for a connection attempt that failed.
(PR #5770)
Fixed LiveKitTransport leaving its audio track unpublished when a track publish failed while connecting and the connection was retried.
(PR #5770)
Fixed LiveKitTransport not releasing its outgoing audio source when it disconnected.
(PR #5770)
Fixed GladiaSTTService tearing down and reconnecting its websocket when Gladia reports a translation addon failure, and silently ignoring rejected audio chunks. Both are now reported upstream as non-fatal ErrorFrames while transcription continues on the same connection.
(PR #5772)
MCPClient now drops an MCP session that a tool call found dead. The next tool call connects again. A client that was never started, or that its owner closed, still raises on a tool call.
(PR #5778)
Fixed ElevenLabsDialogueTTSService closing its WebSocket with a 1008 policy violation (voices can only be set in the first message), which lost the rest of the turn. A context that had gone quiet was dropped locally while ElevenLabs was still generating for it, so the next sentence of the turn registered that context a second time.
(PR #5795)
Fixed VADProcessor ignoring its speech_activity_period while the user is speaking.
(PR #5805)
Fixed SageMaker connections with a connection setting the endpoint rejects (such as language_hint on flux-general-en) hanging until the Flux connection timeout. DeepgramFluxSageMakerSTTService now reports SageMaker's 424 error through on_connection_error within a second. When the session does time out, the error names the endpoint's CloudWatch log group, which holds the container's error text.
(PR #5807)
AssemblyAISTTService now raises a clear error as soon as prompt is set on a non-U3-Pro model, instead of forwarding it and letting the connection fail server-side.
(PR #5808)
Fixed Gemini 3 rejecting a request with Function call is missing a thought_signature when the context holds function calls Gemini did not produce, such as those made by another LLM before switching to Gemini or ones added by the application. GeminiLLMAdapter now sends Google's documented placeholder signature for those calls.
(PR #5809)
Fixed a realtime service configured with its own tools (such as UltravoxRealtimeLLMService with one_shot_selected_tools) answering a tool call with the missing-function result when the model called the tool before the first context frame reached the service. The handlers of a service's own tools are now registered when the service starts.
(PR #5812)
Fixed UltravoxRealtimeLLMService never reporting an async tool's result when the result arrived while the bot was still speaking. The result was sent as a second client_tool_result, which Ultravox ignores after the placeholder it already received; it is now delivered as user-side text like a deferred result.
(PR #5812)
Fixed SarvamSTTService transcripts arriving with finalized=False, which made the turn analyzer wait out its timeout for more transcript before ending the user turn. Each transcript is now marked finalized, so the turn can end as soon as the transcript arrives.
(PR #5815)
Fixed GoogleLLMService (and GoogleVertexLLMService) failing with a 400 (Requests ending with a model turn are not supported.) on newer Gemini models, such as the default gemini-3.6-flash, when the context ends with an assistant message. This happens when a tool handler speaks filler with TTSSpeakFrame before its result arrives. A minimal "." user turn is now appended to the request in that case, for every model not known to continue a trailing model turn; the stored context is unchanged.
(PR #5821)
MiniMaxHttpTTSService now resolves a regional language variant to its base language's name for language_boost (Language.PT_BR becomes "Portuguese") instead of sending the raw pt-BR code.
(PR #5824)
ElevenLabsRealtimeSTTService now converts its language setting to ElevenLabs' language codes, so a regional variant such as Language.EN_US connects with language_code=eng instead of being rejected with a 1008 invalid_request close. ElevenLabsSTTService also resolves regional variants to their base language's code rather than sending en-US.
(PR #5824)
Fixed the eval harness scoring bot speech from before the user's turn as the reply, or losing the reply behind it, when that speech was transcribed late. A send now waits until everything the bot said is transcribed, and a turn that talks over the bot drops the transcripts of what it was saying, by when their audio began.
(PR #5825)
Fixed an absent: true response expectation failing on a later sentence of the reply an earlier expectation had already matched. It now counts only a reply the bot began after that match.
(PR #5825)

Performance

Reduced per-frame memory work for rolling audio volume tracking by retaining the window in a mutable buffer.
(PR #5811)

Open source →

1.9.011 Sept 2026Deprecated3 sources agree

LiveAvatarNewSessionRequest now defaults to H264 video encoding, matching the LiveAvatar API's own default. LiveAvatar has deprecated VP8 ; pass video…

Added

Added LatencyBreakdown.contributions, a timeline of the user-to-bot interval whose durations sum to the measured latency. It names the time no service reports — VAD silence, turn detection, turn-completion markers and holds, sentence aggregation, and function handlers — and attributes time a setting governs to that setting rather than to the service it runs inside.

The first thing the bot says is measured too, from where StartupTimingObserver stops — the pipeline ready with a client on it — so the two reports meet rather than overlap. Waiting to be asked is usually most of that wait, and no service reports it.

Each part carries a stable key and an owner_kind — service, setting, bot or pipeline — so data keyed on them survives a label being reworded, and the breakdown says what it was measured_from alongside the total_secs it measured. Print it with LatencyBreakdown.turn_contribution_lines(); UserBotLatencyObserver(min_contribution_secs=...) sets how brief a part must be before it rolls into a single pipeline entry.
```
```
0.200s  endpointing wait     [config: VAD stop_secs]
0.125s  transcription        [DeepgramSTTService#0]
0.336s  LLM inference        [OpenAILLMService#0]
0.024s  turn completion      [config: filter_incomplete_user_turns]
0.359s  speech synthesis     [CartesiaTTSService#0]
1.044s  TOTAL
```
```
(PR #5445)
Eval scenario turns can play an audio file as the user instead of synthesizing text. A turn's audio: names a recording, resolved relative to the scenario file, in any format soundfile reads (WAV, MP3, FLAC, OGG, ...); multi-channel audio is downmixed to mono and the file keeps its own sample rate. user: is required alongside it and is what the recording says, so the judge and text_contains still have the turn's input. A scenario whose audio turns all name a file needs no user.speech: block.
(PR #5473)
Added a cancellable_by_llm argument to LLMSwitcher.register_function(), which forwards it to every LLM the switcher fronts. Tools registered through a switcher can now opt into LLM-side cancellation, matching what LLMService.register_function() accepts.
(PR #5474)
Added AssemblyAISyncSTTService, a segmented speech-to-text service backed by AssemblyAI's Sync API: pipeline VAD segments the audio and each segment (up to 120 seconds) is transcribed in one HTTP request, with no streaming session to manage. It takes an aiohttp_session, base_url selects a data-residency endpoint, and Settings supports model, language, prompt, keyterms_prompt, and conversation_context.

Recent user and agent turns are sent automatically as conversation_context on each request, bounded by max_context_turns (0 disables) and max_context_chars; setting conversation_context yourself replaces that buffer. The connection is pre-warmed when the user starts speaking so the transcription request skips the handshake (enable_prewarming).
(PR #5488)
AzureSTTService.Settings gained segmentation_silence_timeout_ms, which sets how much silence (100–5000 ms) Azure allows inside a phrase before it emits a final transcript. Azure's default of 500 ms applies when it's unset.
(PR #5511)
Added on_progress and on_update event handlers to the eval framework. EvalSession and EvalSuite are now BaseObjects, so eval progress is observed the same way as every other Pipecat event. EvalSuite's on_update handlers run synchronously and should return promptly, since an EvalRun is mutated in place as it executes; EvalSession's on_progress handlers run as tasks, and EvalSession.run() waits for them before it returns.
```
  @session.event_handler("on_progress")
  async def on_progress(session, progress):
      print(progress.event_name, progress.status)
```
(PR #5520)
Added a voice_parameters setting to AzureTTSService and AzureHttpTTSService, which passes SSML's parameters attribute on <voice> through to Azure so HD voices can be tuned (temperature, top_p, top_k, cfg_scale, enhancePronunciation).
(PR #5522)
Added profanity_filter and redact to DeepgramFluxSTTService.Settings and DeepgramFluxSageMakerSTTService.Settings, exposing Flux's profanity masking and number redaction.
(PR #5523)
Added a version field to DeepgramSTTService.Settings (and DeepgramSageMakerSTTService.Settings), pinning transcription to a specific Deepgram model version instead of whichever one latest currently resolves to. Previously reachable only as an untyped extra key, which continues to work.
(PR #5527)
Added thinking to DeepSeekLLMService.Settings, which controls DeepSeek's thinking mode: settings=DeepSeekLLMService.Settings(thinking={"type": "disabled"}) turns off the reasoning pass that V4 models run by default.
(PR #5528)
CartesiaTurnsSTTSettings gained turn_start_threshold, turn_eager_end_threshold, turn_end_threshold and turn_end_timeout_ms, Cartesia's turn detection tuning parameters. Unset by default, so the server's own defaults apply.
(PR #5533)
Added a denoiser_config setting to GoogleSTTService, which passes Google's Chirp 3 background-noise removal (denoise_audio, snr_threshold) through to the recognition request.
(PR #5536)
Added a temperature field to HumeTTSService.Settings, passing Hume's sampling temperature through to the synthesis request. Higher values increase variation, lower values increase consistency; when unset, Hume applies its own per-model default.
(PR #5540)
Added a speed setting to DeepgramTTSService and DeepgramHttpTTSService, passing Deepgram's Aura speech-rate multiplier (0.7 to 1.5) through to /v1/speak.
(PR #5552)
Added a no_verbatim setting to ElevenLabsSTTService and ElevenLabsRealtimeSTTService, which asks Scribe to drop filler words, false starts and non-speech sounds from the transcript.
(PR #5555)
Added a hotwords field to WhisperSTTService.Settings, which biases Faster Whisper's transcription towards the words or phrases it names, such as product names or domain jargon.
(PR #5558)
Added a reduce_silence setting to SonioxTTSService, which shortens the pauses between words on models that support silence reduction.
(PR #5567)
Added an include_results setting to SpeechmaticsSTTService, which asks Speechmatics to include word-level results in transcript messages.
(PR #5568)
WebsocketService now accepts reconnect_backoff_min_wait and reconnect_backoff_max_wait to configure the wait between reconnection attempts. Defaults (4s / 10s) preserve the previous hardcoded behavior.
(PR #5586)
Added MetaSTTService, a streaming speech-to-text service using Meta's muse-voice-transcribe-1.0 model over the Muse Voice realtime API. Install it with pipecat-ai[meta], or pick Meta when scaffolding with pipecat init. MetaSTTService.Settings takes language or language_bias to bias recognition across the 25 supported languages, keywords for domain-specific terms, and mode to select the model's own endpointing, client-delimited turns, or diarization.
(PR #5587)
Added universal-3-6-pro as a supported AssemblyAISTTService model. It is universal-3-5-pro upgraded — the same model with the same feature set — and 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. universal-3-5-pro remains supported and stays the default.
(PR #5598)
StartupTimingReport.warmup reports what warming Pipecat's deferred imports cost startup, which no processor accounts for and which is often the largest single contributor to a cascaded bot's start. warmup.duration_secs is how long warming took and warmup.blocking_duration_secs is what it added beyond the slowest processor — zero for a pipeline whose services take longer to connect than warming takes to load. Observers can also handle the new BaseObserver.on_startup_warmup() event directly.
(PR #5600)
StartupTimingReport now accounts for the whole of a pipeline's startup, so a slow start can be read off the report rather than inferred. setup_phase_secs and start_phase_secs split the span into the two phases that compose differently: processors are set up concurrently, so the longest single piece of work decides that phase, while the StartFrame reaches them one at a time, so what each spends on it adds up. ProcessorStartupTiming.start_duration_secs reports the latter per processor alongside the existing setup_duration_secs.
(PR #5600)
Added language_codes and timestamps settings to AssemblyAISyncSTTService, aligning it with the Sync API's config surface. language_codes (e.g. [Language.EN, Language.ES]) declares the audio's languages for multilingual or code-switching audio, and is bound in preference to the single language when both are set; regional variants resolve to their base code and duplicates are dropped, preserving declaration order. timestamps requests per-word start/end timings on the transcription result, and is unset by default so the API default of False applies.
(PR #5603)
Added ServiceMetricsObserver, which reports each metric a service publishes as a record rather than a log line. on_service_latency carries time to first byte, first audio and first answer token, with the parts each decomposes into; on_service_usage carries speech-to-text audio seconds, text-to-speech characters, and the token counts an LLM reports. One record per metric, never summed, so a consumer groups them by turn, session or model.
(PR #5607)
Added SpeakingObserver, which reports a conversation's speaking lifecycle through on_speech_event: user_speech_started / user_speech_stopped as the detector heard them, user_turn_started / user_turn_stopped as the turn strategy ruled on them, bot_speech_started / bot_speech_stopped, and interruption. Speech is timed to where it began and ended rather than to where the detector confirmed it, and a moment that closes a stretch of speech names its started_at, so an interval reads whole from one record.
(PR #5612)
Added ErrorObserver, which reports every error a pipeline raises through on_error as an ErrorEvent: the message, the category it was attributed to, the exception_type behind it, the processor that raised it, and whether that processor can still do its job (processor_usable). Errors are read where they are raised, so ones a processor answers for itself — a ServiceSwitcher failing over, or a service it holds in reserve — are reported too; PipelineWorker's on_pipeline_error sees only the errors that reach the top of the pipeline.
(PR #5618)
Added FunctionCallObserver, which reports each function call through on_function_call_event as a FunctionCallEvent: function_call_started when the LLM asks for it, function_call_in_progress when it begins running, and one of function_call_completed, function_call_failed, function_call_timed_out or function_call_cancelled when it settles. Each moment names the one before it, so the wait to run and the time spent running each read from one record. Arguments travel by default and results do not; include_arguments and include_results change either.
(PR #5622)
DeepgramFluxSTTService, DeepgramFluxSageMakerSTTService and CartesiaTurnsSTTService push an EagerTranscriptionFrame when they predict an end of turn and an EagerEndOfTurnCancelFrame when they withdraw it, which EagerUserTurnStrategies acts on. Both frames are emitted only while enable_eager_end_of_turn is on.
(PR #5625)
Added support for answering an eager end of turn: a response is generated from a service's predicted end of turn and held until the turn is confirmed, so the gap before the committed end of turn is spent generating rather than waiting. Turn it on with enable_eager_end_of_turn=True on DeepgramFluxSTTService, DeepgramFluxSageMakerSTTService or CartesiaTurnsSTTService, which then recommend EagerUserTurnStrategies in place of ExternalUserTurnStrategies. The LLM service holds the response, so no extra processor is needed. The response is discarded if the user resumes speaking or if the committed transcript differs from the predicted one, per the strategy's match_policy (NormalizedMatch by default, which ignores the capitalization and punctuation services commonly add when committing a transcript, or ExactMatch to require the two to be identical). An unconfirmed turn never reaches the user or the LLM context, and tool calls are never executed for one.
(PR #5625)
Added FlowConfig, which describes a Pipecat Flows conversation as data so a bot can load its flow at runtime. A config names the initial node and, for each node, its messages, the tools it offers, and where each tool leads. Load it with FlowConfig.from_file(), from_yaml(), or from_json(), construct a Flow from it and the module holding your direct functions and action handlers, or a list of modules, and hand that to FlowManager.

Tools stay in Python and return (result, TRANSITION_IN_YAML); a function
that only moves the conversation is written in the config alone, as a
transition_only entry with a description. The config picks the next node,
by name or with a branch table keyed on a field of the result, where case
keys may be strings, booleans, or numbers. Constructing the Flow reports
every tool, handler, or variable the config names but cannot resolve,
together, as a FlowReferenceError. Actions use the built-in tts_say,
end_conversation, and function types, or a custom type registered in code
or given a handler name in the config. Prompts may use {{ key }}
placeholders, which FlowManager fills from its state on entering the
node, and a YAML config can keep long prompts in their own files with
!include path, resolved relative to the config. The format's JSON Schema is
kept in the repository at src/pipecat/flows/flow_config.schema.json for
editors and other tools to vendor.
```
initial_node: greet
nodes:
  greet:
    role_message: You take food orders for {{ restaurant_name }}.
    task_messages:
      - role: developer
        content: Greet the caller and ask: pizza or sushi?
    functions:
      - name: choose_pizza
        transition_only: true
        description: The caller wants pizza.
        transition_to: pizza
  pizza:
    task_messages:
      - role: developer
        content: Take the pizza order.
    functions:
      - name: select_pizza_order
        transition_to:
          field: status
          cases:
            ok: confirm
            unavailable: pizza
```
```
import handlers  # direct functions: select_pizza_order, ...; action
handlers

config = FlowConfig.from_file("flow.yaml")
flow = Flow(config, handlers=handlers)
flow_manager = FlowManager(
    worker=worker,
    llm=llm,
    context_aggregator=context_aggregator,
    global_functions=flow.global_functions,
)
flow_manager.state["restaurant_name"] = "Luigi's"  # fills {{
restaurant_name }}
await flow_manager.initialize(flow.initial_node)
```
The flows examples are grouped by form: examples/flows/yaml/ holds a
folder per YAML flow with bot.py, flow.yaml, and handlers.py
(hello_world, food_ordering, restaurant_reservation, patient_intake,
insurance_quote, and podcast_interview), and examples/flows/python/ holds
the flows that show what needs code. The README says when to write a flow in
each form.
(PR #5628)
Added pipecat.utils.yaml.include_loader(), which returns a yaml.SafeLoader subclass that resolves !include <relative-path> tags against a base directory. Flow configs and eval scenarios use it; any YAML-based configuration can.
(PR #5628)
A suite manifest lists scripted and simulated scenarios side by side under scenarios:, and pipecat eval suite -k simulation (or -k script) runs only one kind.
(PR #5646)
Simulations for pipecat eval: give a scenario a persona: and a goal:
instead of scripted turns, and an LLM plays the caller, holding the whole
conversation with your bot, in text or over synthesized speech, and hanging
up once it has what it came for. A judge then reads the transcript, the bot's
tool calls in place, and decides whether the bot did its job (success:) and
how every reply scored on your metrics:, judged per turn (criterion, with
min_score as the share of turns that must pass) or measured from the run
(measure: turns, duration, words, or latency against a range, or
function_calls against the list of calls the bot should make, [] for
none). pipecat eval run plays one; a suite runs it runs: times and every
run must pass.
```
persona: "Jamie, booking dinner for two tonight."
goal: "Book a table for two at 6 PM, then end the call."
success: "the bot confirmed a reservation for two at 6 PM"
metrics:
  - name: politeness
    criterion: "the reply is courteous, never curt or dismissive"
    min_score: 1
  - measure: latency
    max_value: 5
runs: 3
```
(PR #5646)
RTVIClientTransport and RTVIClientSerializer let a Pipecat pipeline join a bot as an RTVI client. The eval harness now runs on them: its STT, TTS, persona, and recording are stages of a pipeline of its own, and each suite run gets its own process.
(PR #5646)
Added require_given() to pipecat.utils.types. It narrows a settings value past NOT_GIVEN and None like assert_given(), and also rejects the empty string, raising ValueError("<what> must be specified") for settings a service cannot run without.
(PR #5670)
A simulation's results tell judge trouble from bot trouble: a turn the judge left unanswered carries verdict: none, a failed metric carries a failure_kind (judge_no, judge_no_verdict, out_of_range, function_calls), and a judge that gives no verdict on the goal makes the run an error with kind judge_no_verdict, kept out of the suite's pass rate. A failed latency measure's reason says what was timed, the first token in text mode or the first spoken sentence in audio mode.
(PR #5678)
A simulation ends as soon as it is going nowhere: silence when neither side does anything for max_silence_s (30 s by default), and an error the moment the harness's own pipeline fails, the persona LLM first among them, instead of waiting out max_duration_s.
(PR #5678)
A simulation's simulator: block is optional: without one the persona runs on the same local Ollama model as the default judge, gemma4:12b, so a simulation needs no API key. The release-eval simulations use it.
(PR #5678)
Added BackendLLMWorker and BackendOutput (pipecat.workers.llm): a worker that runs any LLM service, with its own context and multi-step tool calling, as the backend OpenAILiveLLMService client delegation hands work to. Everything the backend produces comes back as a BackendOutput saying what it is and whether the user may hear it, and transform_output decides that per output, or rewrites the text on its way out — see examples/realtime/realtime-openai-live-client-delegation-spoken-updates.py.
(PR #5688)
Added OpenAILiveLLMService, a speech-to-speech service for the OpenAI Live API (gpt-live-1). The live model is full-duplex — it listens and speaks at the same time and handles being interrupted itself — and delegates search, reasoning and tool use to a backend model while the conversation continues. Both delegation modes are supported: OpenAILiveLLMService.ResponsesDelegation lets OpenAI host the backend (Responses API) model, with its function calls executed by the pipeline's registered tool handlers; OpenAILiveLLMService.ClientDelegation runs any Pipecat LLM service as the backend through a BackendLLMWorker. See examples/realtime/realtime-openai-live-responses-delegation.py and examples/realtime/realtime-openai-live-client-delegation.py, plus examples/realtime/realtime-openai-live-client-delegation-spoken-updates.py for a backend that chooses what the user hears.
(PR #5688)

Changed

⚠️ Behavior change: DeepgramSTTService and DeepgramSageMakerSTTService no longer send profanity_filter, so Deepgram's own default of off applies. Deepgram's filter rewrites the words it matches rather than tagging them, so a false positive silently corrupts a transcript that may be stored, analyzed, or sent downstream. To keep the previous behavior, pass settings=DeepgramSTTService.Settings(profanity_filter=True).
(PR #5462)
CartesiaTTSService and CartesiaHttpTTSService now default to sonic-3.6, Cartesia's current Sonic model. Pass model="sonic-3.5" to stay on the previous default.
(PR #5472)
soundfile is now a required dependency instead of an optional extra, so SoundfileMixer and the eval harness can load audio files with no extra install step. The pipecat-ai[soundfile] extra still resolves and is now a no-op, so existing installs keep working unchanged.
(PR #5473)
text_contains on a user_transcription eval expectation now accumulates the STT's final transcription segments within the turn before checking, so an STT that finalizes an utterance in pieces can still satisfy a phrase. Substring checks also ignore whitespace differences between the expected text and the event's text.
(PR #5473)
Raised the pipecat-ai[cli] Context Hub floor to pipecat-ai-context-hub 0.6.0, which registers its MCP server with Claude Code for every directory rather than only the one install ran in. A project scaffolded after the hub was set up now has the server without re-running install.
(PR #5480)
Bots scaffolded by pipecat init build their WorkerRunner before wiring up event handlers, so handlers can end the session with await runner.cancel(). The runner takes handle_sigint=runner_args.handle_sigint, letting a bot run directly install its own signal handler while a hosted one leaves signals to its host.
(PR #5482)
The AssemblyAI STT examples and docs now name universal-3-5-pro, AssemblyAI's current Universal-3 Pro streaming model. The u3-rt-pro names are legacy aliases AssemblyAI redirects to it and are still accepted by AssemblyAISTTService.
(PR #5507)
AWSTranscribeSTTService now maps 46 more AWS Transcribe streaming languages, including Turkish, Hungarian, Tamil, Telugu, Swahili, Mexican Spanish and Welsh. Passing one of these as a Language enum previously sent a bare base code such as tr, which AWS Transcribe rejects.
(PR #5509)
AWSPollyTTSService now maps Language.EN_IE and Language.EN_SG to Polly's en-IE and en-SG locales, so Irish and Singaporean English no longer log an unverified-language warning.
(PR #5510)
Raised the pipecat-ai[cli] Context Hub floor to pipecat-ai-context-hub 0.7.0, whose refresh indexes the framework at its newest release tag rather than tracking its default branch. Hub answers stop mixing in unreleased APIs stamped with the previous release's number, which version_compatibility reported as compatible to anyone running that release.

Pass --framework-version head to index the default branch — the right choice when developing Pipecat itself or building against unreleased code. A --framework-version tag that does not exist now fails the refresh instead of logging a warning and exiting 0.
(PR #5513)
SarvamTTSService and SarvamHttpTTSService now default to bulbul:v3, Sarvam's current TTS model. Sarvam's API no longer serves bulbul:v2 and rejects requests for it, so the previous default could not synthesize. The v3 defaults come with it: shubh as the speaker and a 24000 Hz sample rate.
(PR #5517)
SmallestTTSService now maps all 31 languages the Waves Lightning API accepts, and no longer maps Language.HE, which the API rejects.
(PR #5519)
FireworksLLMService now defaults to accounts/fireworks/models/nemotron-3-ultra-nvfp4. Fireworks no longer serves the previous default, accounts/fireworks/models/firefunction-v2, on its serverless endpoint, so requests using it returned a 404.
(PR #5524)
CartesiaTTSService and CartesiaHttpTTSService now map Language.OR (Odia) and Language.UR (Urdu), the two languages Sonic 3.6 added.
(PR #5532)
Added a turn_coverage setting to GeminiLiveLLMService.Settings, which selects how much of the realtime input stream a user turn covers (for example TURN_INCLUDES_ONLY_ACTIVITY to keep video frames outside detected audio activity out of the turn).
(PR #5535)
GroqTTSService now supports a variety of sample rates, defaulting to 24khz.
(PR #5538)
LiveAvatarNewSessionRequest now defaults to H264 video encoding, matching the LiveAvatar API's own default. LiveAvatar has deprecated VP8; pass video_settings=VideoSettings(encoding=VideoEncoding.VP8) to keep the previous encoding while it lasts.
(PR #5539)
KokoroTTSService.Settings has a speed field, kokoro-onnx's speech rate multiplier (0.5 to 2.0, default 1.0), settable at construction and at runtime.
(PR #5543)
OpenAIResponsesLLMService's service_tier documentation now lists OpenAI's full set of tiers, including fast — the low-latency tier that replaced the priority name.
(PR #5548)
The OpenAI server-side turn detection example now uses gpt-transcribe, replacing gpt-4o-transcribe, which shuts down on 2027-02-26.
(PR #5549)
OpenAISTTService now defaults to gpt-transcribe, OpenAI's replacement for gpt-4o-transcribe, which shuts down on 2027-02-26. With include_prob_metrics=True, GPT transcription models now request logprobs alongside a json response; only Whisper models use verbose_json.
(PR #5549)
PocketTTSService now loads pocket-tts's distilled models for German, Spanish, Italian and Portuguese instead of the 24-layer variants, which are too slow to synthesize in real time on a typical CPU.
(PR #5553)
FalImageGenService.Settings now accepts negative_prompt and guidance_scale, passed through to Fal's image generation request when set.
(PR #5556)
MoonshineSTTService now maps Language.DE (German) and Language.TL / Language.FIL (Tagalog/Filipino), the languages Moonshine added to its model catalogue.
(PR #5557)
TogetherLLMService now defaults to zai-org/GLM-5.2. The previous default, zai-org/GLM-5.1, has been retired from Together's serverless inference and returns model_not_available. Pass settings=TogetherLLMService.Settings(model=...) to choose a different model.
(PR #5570)
NeuphonicTTSService and NeuphonicHttpTTSService now accept temperature in their settings, passing Neuphonic's synthesis randomness control (0.0–1.0) through to the API. Left unset, Neuphonic's own default applies.
(PR #5578)
NvidiaLLMService now defaults to nvidia/nemotron-3-super-120b-a12b. The previous default, nvidia/nemotron-3-nano-30b-a3b, reached end of life on NVIDIA's NIM cloud endpoint and now returns HTTP 410.a
(PR #5579)
⚠️ Behavior change: A word-timestamp event that matches nothing left to speak is now dropped instead of being emitted as a TTSTextFrame. Passing it through put text the LLM never wrote into the conversation context. Nothing is lost: the text that word should have covered is carried by the word that puts the sentence back in step, or emitted when the audio context ends.
(PR #5585)
⚠️ Mem0MemoryService now adds retrieved memories with the "developer" role instead of "system" when add_as_system_message is set. Memories are extra guidance for the model, not a system prompt. Services without a "developer" role receive them as "user" — the same treatment non-OpenAI services already gave the "system" message.
(PR #5596)
SegmentedSTTService now appends 0.5 s of silence to each speech segment before transcribing it, so the model hears the end of speech and finishes the last word instead of dropping or garbling it. This applies to every service built on it, such as WhisperSTTService, OpenAISTTService, GroqSTTService and MoonshineSTTService. The padding is submitted to the provider, so it counts toward STT usage metrics. Pass trailing_silence_secs=0 to the service to disable it, or another value to tune it.
(PR #5621)
FunctionCallResultFrame now carries an error field, set when a call ends because its handler raised, so a failed call can be told apart from a successful one. The result is unchanged: it remains the stand-in message the LLM reads.
(PR #5622)
⚠️ Behavior change: the Deepgram Flux services no longer push their eager end-of-turn transcript as an InterimTranscriptionFrame. That transcript is a prediction the service may withdraw, and it now travels as an EagerTranscriptionFrame, emitted only under enable_eager_end_of_turn, so it is no longer surfaced to clients as partial user speech. Flux emits no interim transcriptions as a result; subscribe to on_update for incremental transcript text.
(PR #5625)
The on_user_turn_inference_triggered event emitted by UserTurnController and by BaseUserTurnStopStrategy now carries a UserTurnSpeculation | None as its last argument, set when the inference answers a turn that hasn't ended yet and carrying the turn text it was run against. A custom stop strategy requests one by passing it to trigger_user_turn_inference_triggered(). The events of the same name on LLMUserAggregator and UserTurnProcessor are unchanged, so handlers registered on those keep their signature.
```
```python
@controller.event_handler("on_user_turn_inference_triggered")
async def on_user_turn_inference_triggered(controller, strategy,
```
speculation): ...
```
(PR #5625)
LLMContextFrame carries a speculation flag marking an inference as speculative, so the response it produces is held until the turn it answers is confirmed.
(PR #5625)
The AGENTS.md that pipecat init writes now describes Pipecat Flows as part of Pipecat (pipecat.flows) and directs coding agents to build a new flow as a FlowConfig YAML file plus a Python tools module, reserving Python-built flows for routing or prompts that need code.
(PR #5628)
Eval suite results.jsonl records carry scenario and kind (script or simulation), and a repeated sweep (--repeat) reports pass rates and exits 0.
(PR #5646)
pipecat eval loads the nearest .env (walking up from the working directory, shell variables winning) before running, so a simulation's persona LLM and a hosted judge find their credentials the way the bots do. python-dotenv is now a core dependency.
(PR #5646)
CrusoeLLMService now defaults to openai/gpt-oss-120b. The previous default was zai/GLM-5.2. Pass settings=CrusoeLLMService.Settings(model=...) to choose a different model.
(PR #5656)
The Speechify service now sends the standard Speechify-Caller-Version attribution header (valued from pipecat_version()), replacing the non-standard X-Pipecat-Version header.
(PR #5668)
The runner extra now requires pipecat-ai-prebuilt>=1.1.0. The prebuilt client UI served by the development runner gains a LiveKit option in its transport selector, backed by @pipecat-ai/livekit-transport 1.0.0, and moves to @pipecat-ai/voice-ui-kit 0.14.0.
(PR #5675)
EvalSession.from_scenario() builds the session for a loaded scenario of either kind, an EvalScriptSession for a scripted scenario or an EvalSimulationSession for a simulation, and EvalSession is the base class of both. Constructing EvalSession(...) directly is no longer supported; use the kind's own class.
(PR #5676)
EvalClientParams holds only what the scenario asks of the bot, and EvalClient takes it together with the run's EvalSessionParams and the services. EvalClient.for_scenario() and EvalClient.for_simulation() are gone; the sessions construct the client.
(PR #5676)
The eval sessions and EvalSuite.run() take how a run behaves as one EvalSessionParams (connect_timeout_s, default_timeout_ms, record_path, cache_dir, use_cache, stop_bot, trigger_disconnect), and the services a run uses (judge, user_tts, bot_stt, and a simulation's persona_llm) as keyword arguments on both from_scenario() and the constructors.
(PR #5676)
Cartesia word timestamps for Chinese and Japanese now keep one token per reported entry, each with its own start time, instead of joining a whole word_timestamps message into a single token.
(PR #5679)
PiperTTSService downloads voice models to ~/.cache/pipecat/piper by default, the same cache location KokoroTTSService uses, instead of the current working directory. Pass download_dir to keep models elsewhere; a model already downloaded into a working directory is fetched again into the cache on first use.
(PR #5687)

Deprecated

Deprecated LatencyBreakdown.chronological_events(). Use LatencyBreakdown.turn_contribution_lines() instead, which names every part of the interval rather than the services that happened to report a metric, and whose lines sum to the measured latency. Will be removed in 2.0.0.
(PR #5445)
Deprecated SarvamTTSModel.BULBUL_V2. Sarvam's API rejects bulbul:v2 on both the REST and the WebSocket endpoint, so it cannot synthesize; use SarvamTTSModel.BULBUL_V3. Constructing SarvamTTSService or SarvamHttpTTSService with it still resolves a model config and now emits a DeprecationWarning. It will be removed in 2.0.0.
(PR #5517)
Deprecated the on_progress parameter of EvalSession() and EvalSession.from_scenario(), and the on_update parameter of EvalSuite.run(). Use the event handlers of the same name instead. Passing a callback still works, keeps its existing signature, and emits a DeprecationWarning. They will be removed in 2.0.0.
(PR #5520)
Setting the system prompt as a "system" message at the start of LLMContext is deprecated and will stop working in 2.0.0. Set it on the LLM service instead:
```
```python
# Before
context = LLMContext([{"role": "system", "content": "Be helpful."}])

# After
llm = OpenAILLMService(api_key=..., system_instruction="Be helpful.")
context = LLMContext()
```
```
To change the prompt at runtime, push an LLMUpdateSettingsFrame carrying LLMSettings(system_instruction=...) instead of rewriting the first context message.
(PR #5596)
EvalScenario, EvalResult, EvalTurn, EvalTurnResult, and EvalTurnProgress are now EvalScriptScenario, EvalScriptResult, EvalScriptTurn, EvalScriptTurnResult, and EvalScriptTurnProgress, the scripted kind of scenario. The old names remain as aliases until 2.0.0.
(PR #5646)
pipecat.evals.harness moved to pipecat.evals.script_session. The old module path remains until 2.0.0.
(PR #5646)
RTVIEvalSerializer is now EvalSerializer, the serializer of the bot's EvalTransport; its client-side counterpart is EvalClientSerializer. The old name remains as an alias until 2.0.0.
(PR #5646)
Deprecated the connect_timeout_s, default_timeout_ms, record_path, cache_dir, use_cache, stop_bot, and trigger_disconnect keyword arguments of EvalSession.from_scenario() and EvalScriptSession.from_scenario(), and the use_cache and default_timeout_ms keyword arguments of EvalSuite.run(). Pass params=EvalSessionParams(...) instead. The old arguments still work, override the params field of the same name, and emit a DeprecationWarning. They will be removed in 2.0.0. EvalSimulationSession.from_scenario() is new and takes params only.
(PR #5676)
Deprecated service: openai in a scenario's judge.eval: and simulator: blocks and service: cartesia in its user.speech: block, along with the openai_service() and cartesia_service() builders behind them. The built-in names are the local, keyless services (Ollama, Kokoro, Moonshine, Whisper); any other provider is a factory:, a dotted path to a callable that takes the block and returns the service, documented with examples in pipecat.evals.services. The old names still work and emit a DeprecationWarning; they will be removed in 2.0.0.
(PR #5678)

Removed

Removed SEND_CHUNK_MS from pipecat.evals.harness. The harness now sends the user's audio through its own output transport, which paces it like any Pipecat transport does.\n
(PR #5646)

Fixed

Fixed AsyncAITTSService ignoring runtime model, voice and language updates. Those three are sent only in the websocket init message and never repeated per utterance, so a change to any of them now opens a new session instead of being stored and warned about.
(PR #5464)
Deepgram Flux STT now applies a model or numerals update by reconnecting, since Flux reads those only from the connection URL. The reconnect waits until the user stops speaking, and settings Flux accepts on a live connection (keyterm, eot_threshold, eager_eot_threshold, eot_timeout_ms, language_hints) are still sent without dropping it.
(PR #5477)
A Deepgram Flux FatalError now reports the code and description Flux sends, instead of a generic Unknown error.
(PR #5477)
Deepgram Flux STT now fails instead of hanging when a connection cannot be established. A connection setting the endpoint rejects leaves the SageMaker session opening forever and produces no confirmation on either transport, so both the session handshake and the wait for Flux to confirm it are now bounded and reported through on_connection_error.
(PR #5498)
Deepgram Flux STT now marks itself unusable once it can no longer transcribe — after a fatal error from Flux, or a connection that never establishes — so the failure reaches the pipeline instead of the service staying connected and silent.
(PR #5498)
Fixed GladiaSTTService dropping Gladia's stop_recording message on a graceful end. The websocket closed before the message was sent, so sessions ended by dropping the connection rather than closing cleanly, and the service's on_disconnected event fired twice.
(PR #5499)
Fixed the pipeline idle timeout cancelling a live session while the user is still talking, on pipelines that get their turns from a provider instead of local VAD. idle_timeout_frames now counts UserStartedSpeakingFrame, TranscriptionFrame and InterimTranscriptionFrame as user activity, alongside BotSpeakingFrame and the VAD-driven UserSpeakingFrame.
(PR #5502)
Corrected the Nova Sonic region lists in AWSNovaSonicLLMService's docstring and example: eu-north-1 is a supported region for both Nova 2 Sonic and Nova Sonic.
(PR #5508)
Fixed the Vonage-Google package conflict so it no longer prevents GeminiSTTService usage in environments where the Vonage connector can't be installed anyway, such as development environments on macOS. The repo lockfile resolved google-genai to 2.8.0 — below the 2.9.0 the service needs — for every Python 3.13 environment, to satisfy vonage-video-connector's pydantic<2.12 pin, though the connector is installable only on Python 3.13 Linux. That environment is now resolved apart from every other. Installs from PyPI were never affected.
(PR #5512)
Fixed SarvamHttpTTSService ignoring its sample_rate. It named the field sample_rate in the request body where Sarvam's REST API reads speech_sample_rate, so synthesis always came back at the model's default rate while the audio frames were tagged with the requested one — audio played back at the wrong speed for any other rate.
(PR #5517)
Fixed the Fireworks function-calling example, which named accounts/fireworks/models/gpt-oss-20b. Fireworks retired that model from its serverless endpoint, so the example now uses accounts/fireworks/models/nemotron-3-ultra-nvfp4.
(PR #5524)
Fixed the examples/flows/llm_switching.py Bedrock service, which pinned a Claude model Bedrock has retired and a latency mode Bedrock does not offer for its replacement. It now uses us.anthropic.claude-haiku-4-5-20251001-v1:0.
(PR #5525)
Corrected GeminiTTSService's documented model names: they differ between the Cloud Text-to-Speech and Gemini API backends, and the docstrings named only Cloud Text-to-Speech's gemini-2.5-flash-tts / gemini-2.5-pro-tts, which the Gemini API rejects with a 404.
(PR #5526)
Fixed Deepgram STT settings passed through extra being discarded when the key matched a declared settings field (for example extra={"diarize": True}). Such keys are now promoted to their field and sent to Deepgram.
(PR #5527)
Fixed GeminiTTSService ignoring its language setting on the Gemini API (GenAI) backend. The configured language is now sent as the speech config's language_code, as it already is on the Cloud Text-to-Speech backend.
(PR #5537)
LmntTTSService now maps all 31 languages LMNT's Blizzard model speaks, adding Assamese, Bengali, Czech, Danish, Finnish, Malayalam, Marathi, Slovak, Tamil and Telugu. These languages already reached the API through the base-code fallback, which logged a "not verified" warning for each of them.
(PR #5544)
Fixed an OpenAISTTService error where include_prob_metrics=True sent a request that diarization models reject. Those models report no per-token or per-segment probabilities, so the service now transcribes without requesting them and warns once.
(PR #5549)
Fixed the examples/voice/voice-pockettts.py import of PocketTTSService, which named a module that does not exist.
(PR #5553)
Fixed an issue where ElevenLabsRealtimeSTTService ignored the realtime API's quota_exceeded, unaccepted_terms and invalid_request messages instead of reporting them as errors.
(PR #5555)
Fixed the PCM sample rates GrokRealtimeLLMService accepts: the Grok Voice Agent API's 22050 Hz option is now allowed, and 21050 Hz — which the API rejects — is not.
(PR #5560)
Fixed the Rime TTS services' language mapping: Arabic, Italian, Japanese and Portuguese — all served by Rime's coda model — now resolve to Rime's language codes, and a region-qualified language such as Language.EN_US falls back to a base code Rime accepts instead of a BCP-47 tag that made Rime close the connection without synthesizing.
(PR #5564)
SambaNovaLLMService now sends the frequency_penalty, presence_penalty and seed settings, which SambaNova's chat completions API accepts. They were previously dropped from the request.
(PR #5565)
Fixed a SpeechmaticsSTTService error where setting split_sentences raised ValueError: "VoiceAgentConfig" object has no field "split_sentences" and the service failed to construct.
(PR #5568)
Fixed SpeechmaticsTTSService reporting the speechmatics-rt SDK version, rather than Pipecat's, in the sm-app tag it sends to Speechmatics. The tag now carries the Pipecat version, matching SpeechmaticsSTTService.
(PR #5569)
Fixed the NVIDIA examples update-settings/llm/llm-nvidia.py and voice/voice-nvidia-sagemaker.py, which requested Llama models that NVIDIA's NIM cloud endpoint no longer serves.
(PR #5579)
Corrected the model docstring in OneShotInputParams for UltravoxRealtimeLLMService: the field defaults to None, which lets Ultravox pick its current default model, rather than to the legacy fixie-ai/ultravox alias.
(PR #5582)
Fixed word-level tracking stalling for the rest of a sentence when a TTS service reports a word-timestamp event that doesn't match what is being spoken. The next correctly reported word is now matched past the gap and carries the text that was never reported, so TTSTextFrames and RTVI progress keep pace with the audio. Previously the sentence stayed stuck at the first unmatched word, every word after it went unmatched too, and the remainder arrived as a single frame once the audio context ended.
(PR #5585)
Fixed NvidiaLLMService emitting reasoning frames before TTFB metrics, so observers receive TTFB before reasoning output for structured reasoning fields and inline <think> blocks.
(PR #5589)
The LLM usage debug log now includes cache creation tokens alongside cache reads.
(PR #5602)
OpenAI-compatible LLM services now report prompt-cache write tokens as LLMTokenUsage.cache_creation_input_tokens, so usage-based cost tracking accounts for them. Covers BaseOpenAILLMService and the services inheriting it, OpenAIResponsesLLMService, and SambaNovaLLMService. Providers that do not report the count leave the field unset.
(PR #5602)
OutputAudioRawFrames pushed from upstream of an STT servi

Note truncated.

Open source →