{
  "id": 10247512,
  "title": "Barge-in for eight AI voices: what it took to let people interrupt a live AI group call",
  "url": "https://urgent.news/2026/09/27/barge-in-for-eight-ai-voices-what-it-took-to-let-people-interrupt-a",
  "topic": "ai",
  "section": "AI",
  "published": "2026-09-27T16:34:37.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/tashthemaker/barge-in-for-eight-ai-voices-what-it-took-to-let-people-interrupt-a-live-ai-group-call-4ll"
  },
  "original_language": "en",
  "account": "Building an AI Group Call, a mobile app that lets users engage in a live voice call with up to eight AI voices, posed unique challenges. The primary hurdle was enabling human interruption among the AI participants. In a one-on-one voice assistant, this feature is largely taken care of, but in a group setting, a single lapse can lead to confusion and derail the conversation.\n\nThe solution involved several key components. Each call consisted of a transcription session for the user's microphone and a real-time speech session for each AI participant. These sessions were stateless, with the server sending the shared transcript in every turn instruction, ensuring all agents shared context without sharing a session. A director model then decided who spoke next and in what order, defaulting to a round-robin approach.\n\nThe conductor owned the floor, allowing only one voice to speak at a time. My first attempt at detecting speaking involved using transcription events, but this proved ineffective as the transcription lagged speech by seconds and lacked server-side voice activity detection. To improve interruption detection, I developed my own energy-based voice activity detection (VAD) on the raw PCM audio, comparing it to an adaptive noise floor. This method held the agent's audio within about 150 ms of a likely voice and confirmed speech with a short probe.\n\nHolding first and confirming second proved effective, minimizing the risk of interrupting oneself. However, running the call in communication mode further reduced echo issues, but with multiple voices involved, additional echo protection was necessary. This included platform voice processing, forced speakerphone, and layered echo protection, all tested independently. Additionally, a client gate filtered out clearly quieter audio, while a server VAD acted independently, further refining speech detection.\n\nThe director model, running on every turn, had to be chosen for low latency, as its delay directly affected the audio flow. Benchmarked against various models, the one with the fastest median latency was selected, with a parallel fallback in case of delays. By implementing these strategies, the AI Group Call app allows users to interrupt the conversation at any moment, with the room going silent and the best-placed AI answering promptly. This setup creates a more natural interaction akin to a live LLM council, enhancing the overall user experience.",
  "summary": "I build AI Group Call , a mobile app where you put two to eight AI voices on one live voice call and talk an idea through with them. A host, a skeptic, a couple of specialists, or models you pick (Claude, GPT, Gemini, Grok) as seats at the table. The hardest part was not getting eight voices to talk. It was letting a human interrupt them. In a one-on-one voice assistant, barge-in is mostly solved…",
  "key_points": [
    "App enables eight AI voices in live call",
    "Director model decides speaking order",
    "Interruption detection improves user experience"
  ],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}