Monitor live voice AI calls and see a health score for every conversation in a dashboard.
Pinpoint calls with long response delays or dead air so you can fix agent performance.
Compare quality across agents on ElevenLabs, Vapi, or Retell to find the best provider.
Drill into a single call to read the transcript, play back audio, and review turn-by-turn timing.
| ayusrjn/benten | aa2246740/ultimate-design | aclark4life/home-depot-crawl | |
|---|---|---|---|
| Stars | 6 | 6 | 6 |
| Language | Python | Python | Python |
| Last pushed | — | — | 2014-08-10 |
| Maintenance | — | — | Dormant |
| Setup difficulty | moderate | easy | moderate |
| Complexity | 4/5 | 2/5 | 2/5 |
| Audience | pm founder | vibe coder | developer |
Figures from each repo's GitHub metadata at analysis time.
Requires an API key from at least one supported voice AI provider (ElevenLabs, Vapi, or Retell) plus local infrastructure for caching and background workers.
Benten is an audio evaluation and analytics platform built for Voice AI agents. When a customer calls a voice-based AI assistant, Benten listens in and measures how well the conversation went, covering things like response delays, audio quality, silent gaps, and moments where the caller talks over the agent. The goal is to help teams understand not just what the AI said, but how it sounded and felt to the person on the other end. Unlike standard text evaluation tools that only read transcripts, Benten analyzes the actual acoustic and conversational performance of voice calls. It tracks several specific metrics. Turn latency measures how quickly the agent responds after a person finishes speaking, in milliseconds. Speech quality gets scored on a 1 to 5 scale using a deep learning model called NISQA. Dead air percentage captures unintended silent pauses. User interruptions count how often the caller speaks over the agent. A combined health score from 0 to 100 brings all of these together into one overall quality rating. Benten connects directly to three Voice AI providers: ElevenLabs, Vapi AI, and Retell AI. You enter your API key, and Benten automatically discovers your existing agents and syncs their metadata. Once connected, it ingests call data through webhooks and begins running its audio analysis pipeline. That pipeline standardizes audio files, detects speech activity, separates the two speakers, scores speech quality, and calculates latency and silence metrics. Results push to the dashboard in real time. The platform includes a conversations table where you can see all calls with summary metrics, color-coded quality tags, and re-evaluation controls. A detail drawer lets you inspect individual calls turn by turn, with searchable transcripts and audio playback. There is also an agents dashboard that tracks health scores and flags bottleneck incidents for each voice bot. The system is built as a modular set of services: a React frontend, a Python backend API, background task workers, and supporting infrastructure for caching and messaging. You can start everything with a single script, or set up each component manually. The full README is longer than what was shown.
Benten listens to calls between people and voice AI agents, then scores how well the conversation went, measuring response delays, audio quality, silence, and interruptions, and shows results in a live dashboard.
Mainly Python. The stack also includes Python, React, NISQA.
No license information is provided, so default copyright restrictions apply, you can view the code but usage rights are unclear.
Setup difficulty is rated moderate, with roughly 30min to a first successful run.
Mainly pm founder.
This repo across BitVibe Labs
Don't trust strangers blindly. Verify against the repo.