Pipecat 1.14.0 adds Perplexity Agent and NVIDIA Omni

Pipecat v1.14.0 was published on October 24, 2026, five days after 1.13.0. Most of it is new services: Perplexity's Agent API, NVIDIA's Nemotron Omni models and Zoom's Scribe Live speech-to-text. The parts that need action on upgrade are smaller: the old Perplexity service is deprecated because Perplexity ended the Sonar API it relies on, several provider defaults changed, and workers sharing a bus must be upgraded together.

Release facts

Check before upgrading

• Perplexity. PerplexityLLMService is deprecated; the notes say Perplexity has ended support for the Sonar chat completions API it uses. Its replacement, PerplexityAgentHttpLLMService, defaults to openai/gpt-5.6-luna with web search on, and sends Perplexity's fast preset when neither preset nor reasoning is set. Answers can contain citation markers such as [1], which you may want to strip before TTS.
• Shared-bus workers. Adapter modules moved to pipecat.adapters.llm.* (PR 6331). Old import paths still work with a warning until 2.0.0, but a worker on an earlier version cannot read the tool schemas a 1.14.0 worker sends. Upgrade every worker on a bus at once.
• Changed defaults. GoogleVertexLLMService moves from gemini-3.6-flash to gemini-3.8-flash because Vertex AI retires the older model on November 19, 2026. NvidiaLLMService defaults to nvidia/nemotron-3.5-super-vl-120b-a12b, and SpeechifyHttpTTSService to simba-3.5. Set settings.model to pin the old values.
• Rime. RimeV1TTSService now requires an explicit settings.model; the URL path no longer selects the model.
• Markdown links. strip_markdown now reduces links and images to their label or alt text, so URLs are no longer read aloud. Pass strip_urls=False to keep the old output.

What's new

• NVIDIA Nemotron Omni. NvidiaOmniLLMService takes speech, images and video. In its default audio-input mode it replaces the STT step, transcribing and answering each turn in one request, and still needs a TTS. Read user text from on_user_turn_message_added, because the service switches the user aggregator into realtime mode.
• Zoom Scribe Live. ZoomAISTTService, installed with pipecat-ai[zoom-ai], ends turns with Scribe's own voice activity detection. silence_duration_ms defaults to 350 and diarization is off by default.
• Text transforms on the LLM side. LLMTextProcessor gains text_transforms, so a rewrite (removing emoji, redacting a pattern) applies to the context and the client as well as the speech. TTS text_filters and MarkdownTextFilter are deprecated in favour of this, with removal planned for 2.0.0.
• Other services. NVIDIA STT speaker diarization, Speechmatics 8 kHz audio and per-sentence finals, SmallestLLMService, and Twilio API-key auth (plus non-US1 regions in the dev runner) for auto hang-up.
• Silent replies surface as errors. The OpenAI Responses services push a non-fatal ErrorFrame when a response finishes with no text and no function call, so on_pipeline_error can retry instead of the bot staying silent.

Notable fixes

• Each send through WebsocketService is now capped at 5 seconds. Before, a peer that stopped reading could stall a service and everything after it in the pipeline for minutes. Tune it with ws_send_timeout.
• Deepgram Flux and Cartesia Turns STT reconnect in the background and close a half-finished turn with the transcript received so far, instead of losing what the user said.
• DeepgramFluxTTSService stops retrying a refused connection on every reply, so a configured failover can take over.

Questions

Do I have to migrate off PerplexityLLMService now? The release deprecates it because Perplexity ended the Sonar chat completions API behind it, so plan to move to PerplexityAgentHttpLLMService. pipecat init already scaffolds the new service.

Can I upgrade one Pipecat worker at a time? Not if the workers share a bus. The notes say an older worker can't read the tool schemas a 1.14.0 worker sends.

What Python version does Pipecat 1.14.0 need? PyPI lists >=3.11.

More on the Pipecat hub, the previous Pipecat 1.13.0 launch and the voice agents category. The closest alternative tracked here is LiveKit Agents.