Pipecat and Optimizing AI Voice Agents

pipecat looks to be the cli swiss army knife for tts

Kwindla Hultman Kramer, founder of Pipecat.

The talk is Optimizing AI Voice Agents (June 15, 2025) on Hamel's channel — a guest session from his AI Evals course.

https://www.youtube.com/watch?v=I86dFivLzXY

From Hamel's own video description:

Kwindla is the founder of Pipecat, an open source, vendor neutral framework for voice and multimodal conversational AI. Pipecat is used and supported by teams at NVIDIA, Google DeepMind, OpenAI, AWS, and hundreds of startups and scale-ups.

One clarification worth having: Pipecat isn't a TTS model — it's the orchestration framework for the STT → LLM → TTS pipeline, where TTS is one pluggable stage. So he's the voice-AI person, not a speech-synthesis-model person.

The talk itself covers why voice use cases are out-of-distribution for today's SOTA models, why evaluating voice agents is hard, and a DIY segment on saving conversation data (SQLite + WAV) and vibe-coding scripts to actually look at it. The OpenTelemetry-on-a-voice-agent demo at 13:50.

Code from the session:

https://github.com/kwindla/evals-course-voice