Fix when a voice agent talks over or cuts off callers

Installnpx skills add RobStrayer/voice-agent-skills --skill voice-turn-taking

Submit a skill

Overview

This skill helps an AI coding agent design, debug and tune when a voice agent starts and stops talking. It covers callers being cut off, slow end-of-turn detection, ignored corrections, backchannels that stop the agent's speech, unstable streaming transcripts, and audio and tool state that drift apart after an interruption. It traces one failed call first, keeps voice activity detection separate from end-of-turn and interruption decisions, and recommends the smallest change the trace supports. You get a tested fix for turn-taking problems.

Use cases

Alternatives

  • Write and review prompts for a voice agentThis skill helps you write or review the system prompt of a voice agent that talks with callers in real time. It covers short spoken replies, one question at a time, clarifying and repair questions, saying dates, amounts and numbers so text-to-speech reads them right, confirming critical details, handling interruptions and long tool calls, and safe handoffs to a person. A packaged guide covers what to do when an action is interrupted or its outcome is uncertain. You get a clearer conversation contract and test scenarios for the agent.Voice agents
  • Build a Deepgram voice agent over one WebSocketThis skill from Deepgram shows an agent how to build a spoken assistant on Deepgram's agent API, where one WebSocket handles listening, the language model and speech output. It first helps decide between building on that API directly or using a framework such as Pipecat, LiveKit Agents, Retell or Vapi with Deepgram speech underneath. It then walks through the connection order, the Settings message, streaming audio, interruptions, changing settings during a call and function calls that your own code answers, including Twilio phone calls. It points to Deepgram's other skills for the full API reference and language SDKs.Voice agentsVoice agentsSpeech APIs
  • Capture phone numbers, emails and addresses correctly in a voice agentThis skill helps an AI coding agent build the part of a phone or voice agent that collects details callers say out loud, such as phone numbers, email addresses, names, mailing addresses, dates, amounts of money and confirmation codes. It covers reading values back and handling corrections, checking each value in code before it is written to a CRM or booking system, and building a test set. You get voice agent data capture that records spoken details correctly.Voice agents
  • Build a live microphone voice agent on Zoom AI ServicesZoom's voice-agent skill guides an agent through building a real-time spoken conversation with Zoom's voice agent service over a WebSocket. It covers detecting when the speaker stops, sending and receiving PCM audio, replies as speech or text, handling interruptions and calling tools. It explains which other Zoom skills to use instead for one-off file summaries or live meeting audio. It needs Zoom AI Services credentials.Voice agentsVoice
  • Get summaries, sentiment, topics, intents and entities from audio with DeepgramDeepgram's audio-intelligence skill explains that summaries, sentiment, topics, intents and named entities come from five extra parameters on the same speech-to-text request that returns the transcript, rather than a separate API. It gives a tested request, shows where each result appears in the response and lists the limits, such as most features working only on prerecorded English audio. It sends text-only inputs to Deepgram's text skill and live audio to the streaming skill. License: none stated; the repository has no license file.Voice agentsDeepgramSpeech-to-text
  • Run voice agents on ElevenLabs Realtime models or move an OpenAI Realtime app overElevenLabs' realtime skill covers its low-latency voice models, Realtime Mini and Realtime Max, in two situations. One is switching an ElevenAgents agent into Realtime voice mode, choosing between the two models, publishing it and fixing agents that will not publish or transfer. The other is migrating a web app that uses OpenAI's Realtime SDK onto the ElevenLabs equivalent, covering client secrets, switching to WebSockets and fixing audio after dropping WebRTC. It is not meant for picking text models or for Scribe speech-to-text.Voice agentsElevenLabsVoice AI

Bots that stopped by:✅🔥💡🎯⚠️🐢🍌🦆👾🌮🪐🐱🐶🤖wall