Original title: The Machines Are Calling: Measuring Automated and Synthetic Voices in Unwanted Inbound Calls
Xingyu Shen, Tommy Duong, Muduo Xu, Xiaodong An +6 · September 2026
Robocalls bother millions of people, but nobody knew how many use fake, computer-generated voices instead of simple recordings. Researchers set up a phone trap to find out. They put fake personas run by language models on real U.S. phone numbers to answer calls and record the callers. Next, they analyzed the callers' voices. They used audio fingerprinting to catch repeated recordings, an automated detector to spot synthetic speech, and human listeners to verify the flags. They found that machine-voiced openings make up at least 26.9% of calls. Specifically, 13.8% start with a reused recording, and 13.1% open with new audio labeled as synthetic. Another 9.9% are completely silent after the greeting. Only 0.44% admit they are automated. Computer-generated voices appear most often in lead-generation spam rather than fraud. Finally, older bait numbers get far more of these calls than newer ones do.
Imagine a mail carrier delivering letters to a house with a talking doorbell. Some letters come from the same sender each time, others are written in perfect robotic handwriting. You want to count how many are from machines. The mail carrier tries three clues: whether two letters look identical, whether the handwriting is unnaturally perfect, and whether a neighbor agrees it's fake. In this paper, the researchers set up phone numbers as talking doorbells. They used a recording fingerprint, a synthetic-voice detector, and human listeners to measure how many unwanted call openings are from machines.
This matters because regulators like the FCC recently banned AI-generated voices in calls, but had no data on how common they actually are. This study shows that machine voices are already widespread—especially in lead-generation spam—and that detection tools are unreliable. Consumers face an invisible surge of automated calls that mimic human speech, making enforcement harder.
The study only measures call openings, not full conversations. The synthetic-speech detector is imperfect—identical waveforms are flagged inconsistently 13.6% of the time, and human listeners confirm only 54.4% of its flags. Results come from a single honeypot setup on U.S. numbers over 66 days, so may not generalize. The 11 silent days were excluded, and the pipeline may miss machine-voiced calls that do not open with speech.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
The Machines Are Calling: Measuring Automated and Synthetic Voices in Unwanted Inbound Calls — Xingyu Shen, Tommy Duong, Muduo Xu et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.