Pipecat 1.15.0 adds expressive TTS and turn strategy modes

Pipecat v1.15.0 was published on October 29, 2026, five days after 1.14.0, and reached PyPI the same minute. The headline is expressive conversation, an alpha feature that lets the LLM write delivery and sound tags that the TTS service performs instead of reading aloud. The change most likely to affect an existing bot is quieter: when a UserTurnStrategies.stop list holds more than one strategy, all of them now have to agree before the user's turn ends.

Release facts

Check before upgrading

• Several stop strategies now all have to agree. stop_mode defaults to StrategyMode.ALL, so TurnAnalyzerUserTurnStopStrategy next to SpeechTimeoutUserTurnStopStrategy ends a turn only once both decide. Pass UserTurnStrategiesParams(stop_mode=StrategyMode.FIRST) for the old behaviour. A list with one stop strategy is unchanged.
• Wake phrases start turns on their own. WakePhraseUserTurnStartStrategy should now be the only start strategy. To combine it with something like MinWordsUserTurnStartStrategy, set start_mode=StrategyMode.ALL.
• Custom stop strategies. One that decides from the turn's inference must return True from the new waits_for_inference, or the turn controller treats it as a listener. trigger_user_turn_stopped() is replaced by report_user_turn_stopped(), and deferred() is deprecated with no replacement.
• Deprecated marker setting. Turning on turn-completion markers with only filter_incomplete_user_turns no longer re-prompts a quiet user. Use user_turn_strategies=FilterIncompleteUserTurnStrategies().
• Inception default model. InceptionLLMService moves from mercury-2 to mercury-3. The notes say Inception retires Mercury 2 and 2.5 on November 15, 2026.
• LLMWithBackend imports. The module now exports only the classes an app builds with. Prompt texts moved to pipecat.pipeline.llm_with_backend.prompts, and DELEGATE_TOOL_NAME to .connector.

What's new

• Expressive conversation (alpha). Enable it with PipelineParams(enable_expressive_conversation=True). Tags such as <<excited>>, [[laugh]], [[pause 500]] and ^^word^^ are performed by the TTS service, and clients see the words without tags. Adapters ship for Cartesia, ElevenLabs (v3 and v4 models), Fish Audio S2, MiniMax speech-2.8, Inworld, Soniox, Mistral Voxtral 3, Gemini 3.8 TTS and Azure DragonHD voices. The LLM is taught the tags the active TTS can perform, and an ExpressionPolicy limits which ones it may use. The notes warn the tag syntax and APIs may change between releases.
• Pipecat Tone models. PipecatToneReader reads each user turn for likely emotion and sarcasm. PipecatToneTagger writes expression tags for plain LLM replies. Both download from Hugging Face, run on the CPU and install with pipecat-ai[tone].
• Classifier turn ending. ClassifierUserTurnStrategies lets a classifier such as JevClassifier judge whether the user is done, as an alternative to LLM turn-completion markers.
• Turn strategy modes. UserTurnStrategiesParams sets start_mode/stop_mode (FIRST or ALL) and whether your strategies replace or extend the ones a service recommends.
• New and extended services. AssemblyAITTSService (streaming, word timings, EU endpoint option), reasoning output from AWSBedrockLLMService, Tavus replica emotions, an AICTytoObserver audio quality score, plus settings for NVIDIA TTS, Smallest, Cartesia, Speechify, Together and OpenAI image moderation.
• Evals. Scripted expectations gain text_matches regular expressions, and scenarios can check the expression a bot actually performed.

Notable fixes

• Segmented STT services (Whisper, OpenAI, Groq, Moonshine) no longer end a turn on the user's earlier words when the user speaks again before a segment is transcribed.
• AWSBedrockLLMService no longer sends Amazon Nova's <thinking> text to TTS, so voice bots stop reading their reasoning aloud.
• NvidiaTTSService fetches its config in a thread, so a slow endpoint no longer blocks the event loop at startup.

Questions

Will my bot behave differently after upgrading? Only if a stop list has more than one strategy or you use a wake phrase. In those cases turns can end later than before. Set stop_mode=StrategyMode.FIRST to keep the old timing.

Is expressive conversation ready for production? The release labels it alpha and says the tag syntax and APIs are likely to change. Pin the version if you build on it.

What Python version does Pipecat 1.15.0 need? PyPI lists >=3.11.

More on the Pipecat hub, the previous Pipecat 1.14.0 launch and the voice agents category. For a side-by-side with the closest alternative, read Pipecat vs LiveKit Agents.