hey everyone 👋 so we were building a voice bot for customer calls and the testing part was a complete nightmare. every time we pushed a change, we'd literally call the bot ourselves to see if it still worked. obviously that doesn't scale at all. the main pain points were latency (users hang up if the bot takes too long), hallucinations (bot randomly inventing refund policies), and interruptions (bot completely breaking when someone talks over it). we looked for tools but everything was cloud-based SaaS that wanted our call transcripts on their servers. we handle sensitive customer data so that was a dealbreaker. ended up building decibench it just synthesizes a fake caller, sends audio to your agent, and checks if the response is solid. works with websockets, twilio, elevenlabs, vapi, retell. whole thing runs on your machine, nothing leaves. 🔒 been using it internally for a while now and figured we'd open source it. curious how others here test their voice agents or if everyone's just winging it like we were 😅

DeciBench

pytest for voice AI agents. Local-first & open source.

pytest for voice AI agents. Local-first & open source.