Optimizing AI Voice Agents скачать в хорошем качестве

Optimizing AI Voice Agents

Join the AI Evals Course starting Jan 27, 2026: https://maven.com/parlance-labs/evals... . 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. Kwin has been helping people ship generative AI voice agents to production since the early GPT-4 days. Kwin will talk about why most voice AI use cases tend to be somewhat out-of-distribution relative to how today's SOTA models are trained, specific challenges involved in evaluating voice agent performance, and the recent evolution of ops and evals tools to better support voice AI developers. 35% our evals course: https://bit.ly/evals-ai Slides: https://docs.google.com/presentation/... Code: https://github.com/kwindla/evals-cour... 00:00 Introduction & Speaker Background 01:35 Voice Agent Architecture: The Standard Pipeline 03:15 Standard & Complex Pipeline Code Examples 04:22 Building & Deploying Voice Agents: 3 Key Steps 05:33 What Makes Voice AI Hard? Challenges Overview 08:00 Q&A: Handling Long Contexts in Voice Conversations 12:09 Data Observability & Building Your Own Tools 13:50 Demo: Adding Open Telemetry to a Voice Agent 15:20 Q&A: Unexpected LLM Responses & Voice Data Analysis 15:53 DIY: Saving Voice Conversation Data (SQLite & WAV) 17:21 Vibe-coded "Look at the Data" Scripts 17:58 Demo: Analyzing & Playing Back Conversation Data 20:47 Q&A & Final Thoughts on Voice AI Challenges