Releases · unforkopensource-org/decibench

🎉 Decibench v1.0.0 — First Stable Release

The open testing standard for voice AI agents.

10 built-in evaluators — Latency, WER/CER, Hallucination, Task Completion, Compliance, Interruption, Silence, MOS, STOI, Composite Score
8 connectors — WebSocket, HTTP, Process, ElevenLabs, Twilio Mock, Vapi, Retell, Demo
3 testing modes — Deterministic (free), Semantic (LLM Judge), RAG-Augmented (auto-generate adversarial tests from your docs)
MCP Server — AI coding agents can run and analyze voice tests via Model Context Protocol
Vue Dashboard — Failure workbench, call timelines, evaluation breakdowns
Privacy engine — PII redaction (credit cards, SSNs, phone numbers) before storage
Zero telemetry — 100% local-first, runs completely offline with Ollama + local Whisper

pip install git+https://github.com/unforkopensource-org/decibench.git
decibench run target=demo suite=quick