ICT Innovations Releases tvbench, an Open Source Benchmark for Telephony AI Voice Agents

MULTAN, Pakistan - Sept. 8, 2026 - PRLog -- ICT Innovations has released tvbench, a free and open source benchmark that measures a telephony voice agent the way a caller experiences it. It installs from PyPI with one pip command, runs against any agent that accepts an AudioSocket connection, and is MIT licensed.

Voice agents are usually scored on transcription accuracy and model latency. Neither is what a caller hears. A caller hears whether the words arrive evenly, whether the sentence has holes in it, and how long the agent keeps talking after being interrupted. tvbench measures those three things from the far end of the call.

The tool came out of a bug the company could not see. Its own agent sounded broken on real calls while every log said the audio had been delivered perfectly. It had been. Asterisk hands each frame straight to the channel without buffering, so an agent that writes a whole sentence to the socket at once delivers every byte and the caller still hears only the tail, because the jitter buffer at the far end keeps a few frames and drops the rest. Counting bytes cannot find that. Timing arrivals finds it in one run.

Everything tvbench reports is computed from when audio arrived, never from how much. It answers four questions: how long the caller waited to hear anything, how bunched the frames were, how long the longest silence inside the agent's speech was, and how long the agent carried on after the caller started talking.

A benchmark nobody has tried to break is only an assertion, so the tool ships a reference agent with switchable defects and a script that runs the whole matrix. Switch on bursting and the pacing figure moves by a factor of thousands. Make the agent ignore interruptions and the barge-in number goes from 97 milliseconds to just under five seconds.

"We wrote this because our own agent was failing in a way none of our tools could see," said Tahir Almas, founder of ICT Innovations. "If someone runs it against their agent and thinks our numbers are wrong, the results folder takes pull requests. Disagreement with a reproducible number attached is the whole point."

tvbench is direct about what it cannot do. It says nothing about transcription accuracy, because the caller signal it generates is band-limited noise rather than speech. It needs Linux to pace frames accurately.

Design notes and the full falsification pass: https://ictinnovations.com/caller-side-voice-agent-benchm...

Source code: https://github.com/ictinnovations/telephony-voice-agent-b...