Pipecat 1.16.0 adds diarization speakers and drops soxr

Pipecat v1.16.0 was published on November 12, 2026 (03:45 Istanbul time), two weeks after 1.15.0, and reached PyPI a minute earlier. The notes run to more than 240 entries. Three things matter most: diarization now has its own speaker field on transcription frames, the default install is roughly half the size because resampy and soxr left the core dependencies, and Hume's TTS service is deprecated because Hume shuts its TTS API down on November 13, 2026.

Release facts

Check before upgrading

• Diarization moved to speaker. AssemblyAISTTService, SpeechmaticsSTTService and NvidiaSTTService now put the speaker in frame.speaker and leave user_id as the transport participant id. Code that read the speaker from user_id needs to change.
• Speechmatics prefix gone. With the deprecated InputParams and enable_diarization=True, transcripts no longer start with @{speaker_id}: . Set speaker_format="@{speaker}: {text}" to keep it.
• resampy and soxr are no longer core dependencies. Install pipecat-ai[resampy] or pipecat-ai[soxr] if you still use ResampyResampler or the SOXR resamplers, all now deprecated. The default resamplers are the new SincStreamAudioResampler and SincAudioResampler from the pipecat-resampler package.
• audioop-lts dropped on Python 3.13+. Add it yourself if your own code imports audioop. The local-smart-turn extra also stops installing torchaudio.
• Function call results. A function call that returns "", {} or False now runs the LLM. Return None to finish a call without a reply.
• Hume TTS stops working on November 13, 2026. HumeTTSService is deprecated with no replacement and pipecat init no longer offers it.
• Model defaults moved. SonioxSTTService defaults to stt-rt-v6. GeminiTTSService uses gemini-3.8-flash-lite-tts on the Gemini API, ahead of gemini-3.1-flash-tts-preview retiring there on November 17. The aic extra needs aic-sdk~=3.4.0, which cannot download older model files.
• Expressive conversation imports moved (still alpha). PipecatToneReader and PipecatToneTagger now live under pipecat.expressive.processors, and pipecat.processors.tone is removed. The tone extra is deprecated in favour of expressivity.
• Nested !include paths in eval scenarios and FlowConfig now resolve relative to the file that names them.

What's new

• Speaker diarization across STT services. New support in ElevenLabs, AWS Transcribe and OpenAI diarization models such as gpt-4o-transcribe-diarize; Deepgram, Soniox, xAI, Smallest, Meta and Zoom AI now report the speaker too. A speaker_format setting writes it into the transcript text, and RTVI messages carry it to clients.
• New services. GoogleInteractionsLLMService for the Gemini Interactions API (sent with store=False, so Google keeps no conversation state), a Vertex AI variant, ZoomAIRealtimeLLMService for speech-to-speech with the Zoom AI Voice Agent, RimeSTTService, SarvamHttpSTTService and HeyGen Voice TTS.
• Telephony. A JambonzFrameSerializer for jambonz's listen verb, audio_in_sample_rate/audio_out_sample_rate on every serializer, synchronous G.711 helpers in pipecat.audio.g711, and Genesys AudioHook pause and resume.
• Perception for expressive conversation (alpha). PerceptionReader asks classifiers how a user turn was said. Sources include Pipecat Tone, the new on-CPU Pipecat Voice Reader (tone of voice, arousal, laughter, speech rate), Hume Audio Decisions and Google's Decisions API. Classifiers now accept audio and files through inputs=.
• Soniox echo cancellation. enable_echo_cancellation sends the bot's words as they play so Soniox does not transcribe the bot as the user.
• Evals. EvalRubric judges a conversation the harness did not drive, such as a recorded production call.

Notable fixes

• GoogleSTTService no longer drops the user's last utterance each time it restarts at Google's four-minute streaming limit.
• LiveKitTransport no longer leaks memory queuing audio when audio_in_enabled is False.
• SentenceAggregator and IVRProcessor clear partial text on interruption, so cut-off words no longer start the next reply.
• Wake phrases with punctuation, or Chinese and Japanese wake phrases inside longer speech, now match.

Questions

Will upgrading break my bot? It can if you read diarization speakers from user_id, rely on soxr or resampy being installed, return "" or {} from a function call to stay silent, or use Hume TTS. Check those four first.

Why is the install smaller? resampy, which pulls in numba and llvmlite, is no longer a core dependency. The release says this roughly halves the default install.

What Python version does Pipecat 1.16.0 need? PyPI lists >=3.11.

More on the Pipecat hub, the previous Pipecat 1.15.0 launch and the voice agents category. For other options, read Pipecat alternatives and Pipecat vs LiveKit Agents.