Pipecat v1.12.0 (2026-09-26) adds TTSService.pronunciation_transform_ipa() for reusable word→IPA maps across Cartesia/ElevenLabs/Inworld/Deepgram, client-mode MOQTransport reconnect with session-ending markers, Pipecat Classifiers + LLMClassifier, frame-level interruptible, and response_schema on LLM run_inference. PyPI pipecat-ai 1.12.0 is live.

Key Takeaways

  • ✓Shipped: GitHub v1.12.0 = PyPI pipecat-ai 1.12.0; docs docs.pipecat.ai
  • ✓Pronunciation: pronunciation_transform_ipa maps words to IPA once; vendors emit phoneme/SSML/slash IPA—register last in text_transforms
  • ✓Transport: MOQTransport client reconnect within connection_timeout (default 60s); session-ending distinguishes hangup vs outage; relay_url full URL
  • ✓Eval: Classifiers (YesNo/Choice/Score) + LLMClassifier batch questions; allow_continue=false for final-answer judges
  • ✓Inference: run_inference response_schema enforced by OpenAI/Anthropic/Google; frames interruptible by default; UninterruptibleFrame deprecated
🔗

Project Links & Resources

Direct Access
Direct access to official project resources and documentation
🐙
Pipecat v1.12.0 Release
github.com/pipecat-ai/pipecat/releases/tag/v1.12.0
↗
📖
pipecat-ai 1.12.0 on PyPI
pypi.org/project/pipecat-ai/1.12.0/
↗
📖
Pipecat Docs
docs.pipecat.ai/
↗
📝
Pipecat Homepage
pipecat.ai
↗
🐙
CHANGELOG
github.com/pipecat-ai/pipecat/blob/main/CHANGELOG.md
↗
🔬

In-Depth Technical Analysis

Background Voice agents must bind STT→LLM→TTS, flaky-relay reconnect, and eval judges in one framework. Pain points: TTS guessing drug/brand names, per-vendor pronunciation markup, MOQ drops replaying full RTVI logs, judges limited to yes/continue, and empty user turns killing idle timers. Pipecat v1.12.0 (2026-09-26) targets those engineering gaps. ### Architecture TTSService.pronunciation_transform_ipa() builds word→IPA transforms: Cartesia phoneme blocks, ElevenLabs SSML phonemes (WebSocket + enable_ssml_parsing), Eleven v3/Inworld slash IPA, Deepgram Aura-2 inline objects. TTS sentence aggregation follows Settings.language via sentencex (no NLTK download). MOQTransport client mode redials within connection_timeout (default 60s), keeps broadcast, uses session-ending markers, and accepts relay_url. New Classifiers + LLMClassifier; run_inference response_schema enforced by OpenAI/Anthropic/Google. Frames carry interruptible (default True); UninterruptibleFrame and several APIs deprecated toward 2.0.0. ### Benchmarks No public MOS/WER leaderboard. Track unsupported-word warnings from IPA transforms, reconnect success inside the 60s window, and single-call Classifier batching on your suites. ### Getting started pip install -U pipecat-ai==1.12.0 (moq extra needs moq-rs>=0.4.6). Docs: docs.pipecat.ai. Register pronunciation transforms last. XAISTTService defaults to grok-voice-transcribe-2.0—pin 1.0 if needed. Full notes on the GitHub release.