A benchmark tested 14 production voice agents using Veris AI’s VAmoS Pro Bench by running the same 100 calls to a utility billing line (three runs each) and inserting a simulated TV news show into a quarter of the calls. Overall task completion dropped from 38.7% without TV to 8.6% with TV. Street noise had negligible impact, café chatter reduced completion by ~11 points, and background television reduced it by ~30 points. Nine of the 14 agents completed under 5% of TV calls and three completed none; the strongest performers were GPT-Live 1 (36% completion) and Grok Voice Think Fast 2.0 (34%).
The explanation is technical: automatic speech recognition models transcribe all audible speech and typically lack robust speaker-attribution, so agents pick up and respond to TV audio rather than the human caller. That failure mode - background speech rather than ambient noise - breaks most current cascaded and speech-to-speech architectures. The work links to a full Veris AI write-up, open-source agent harnesses, and VAmoS arXiv papers and leaderboards, and highlights that multi-speaker voice-activity detection remains a critical unsolved problem for deploying reliable voice agents in real-world, TV-filled environments.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
Developers disassemble libraries to report bugs due to delays in AOSP source releases. A recent issue reveals that on Android 17 builds, wcscspn() and wcspbrk() functions behave incorrectly with certain wchar_t characters. (github.com)
Anthropic has raised concerns about a potential AI threat that does not currently exist to evoke fears of an AI apocalypse. Critics argue that these warnings are speculative and may be used to regulate or restrict AI development. (theregister.com)
Experts warn parents to prepare children for the potential trauma of witnessing a livestreamed execution. The advice highlights concerns about the increasing possibility of violent events being broadcast online. (apnews.com)
Chinese hackers have targeted US government agencies including the DOJ, NASA, the Federal Reserve, and the Senate. The US government has publicly attributed the cyberattacks to Chinese state-sponsored actors. (usatoday.com)
Scientists warn that AI security failures highlight the need for better containment and alignment practices to prevent damaging incidents. They recommend immediate adoption of high-assurance sandboxing and the creation of an international body to standardize safety protocols and ensure transparency. (ai-security-letter.org)
The Department of Justice arrested two individuals and seized 187 devices in a crackdown on a North Korean-operated laptop farm in Texas. The operation targeted illegal activities linked to North Korea's cyber activities. (reuters.com)
Today's best Hacker News stories, summarized and screenshotted, one email a day.