Pipecat 1.11.0: Resampler with Manual Reset and Function Call Evaluation in Scenarios

On September 18, 2026, Pipecat 1.11.0 was released — an open source Python framework for real-time voice and multimodal agents. Since the last review, two releases have come out: 1.10.0 and 1.11.0.

What's new

• Manual control of audio stream boundaries. BaseAudioResampler gained flush() and reset(): flush() returns audio that the resampler is still holding for a finished stream, reset() discards it for an abandoned one. Previously, boundaries were determined only by the inactivity timeout of SOXRStreamAudioResampler — now the calling code decides where the stream ended.
• Evaluation of function calls in scenarios. The scripted eval for function_call gained the eval:: directive — each matched call is evaluated by name and arguments with a separate judge prompt. A rejected call ends the turn with kind judge_no. This makes it possible to check arguments that cannot be matched verbatim.
• Configuring the stability of AWS Transcribe partial results. AWSTranscribeSTTService.Settings gained partial_results_stability — a choice of "high", "medium" or "low" for the stability of partial recognition results. The default remains "high".
• Effects for Azure TTS. AzureTTSService and AzureHttpTTSService gained the effect setting, which passes an Azure audio effect processor (eq_car, eq_telecomhp8k) into the synthesis request.
• Controlling reasoning effort in Groq. GroqLLMService.Settings gained reasoning_effort — the values depend on the model: "low", "medium", "high" for GPT-OSS and qwen/qwen3.8-27b, "none" and "default" for Qwen models.

Release 1.10.0 (September 12) also introduced: a continuation API for Smallest TTS (text fragments within a single LLM turn are merged into one continuous generation), the audio_out_queue_size_ms setting for LiveKit, server-side turn boundary detection in GradiumSTTService, and a switch to the openai 3 SDK with HTTP clients on httpx2.

What the platform does

Pipecat is a framework for real-time voice and multimodal agents: you can build a single agent or an entire system where specialists hand tasks to each other, work in parallel and coordinate through a shared bus — locally or distributed. The platform integrates speech recognition, speech synthesis and conversation processing, and supports voice, video and images. For debugging it offers OpenTelemetry and Sentry, and for deployment a CLI with monitoring and production deployment.

Your own model instead of the cloud

The documentation indicates that your own model is connected via Ollama. The framework is installed via pip install pipecat-ai, and the license is BSD-2-Clause.

How it differs from its neighbors

A comparison with CrewAI, Agno and AgentScope shows that Pipecat has no features that are claimed only for it. At the same time, Pipecat's documentation does not mention its own tools and functions, MCP support, the A2A protocol, agent memory, a knowledge base (RAG) or human approval — all of these are claimed by most of its neighbors. Pipecat stands out for its focus on real-time voice and multimodal scenarios, whereas its neighbors are more oriented toward orchestrating autonomous agents and working with tools.

It is worth taking a closer look at the update for those who already build voice agents on Pipecat and want more precise control over audio stream boundaries or to check function call arguments in scenarios. You can follow new releases of agent platforms in the CPU3D news feed.