Rob Lesi AI voice auditing is effective, but human audit is still better

　　The large-scale game creation platform "Robu Lesi" has used machine learning AI technology to carry out voice chat review work for a year. Given that the platform has a large group of children’s users, the company always regards audit work as a top priority. At the GDC 2024 game developer conference held recently, Kiran Bhat, senior technical director of Rob Less and Hannes Heikinheimo, head of voice security, revealed that despite the significant results of AI audits, manual audits still have irreplaceable advantages in some cases.

　　Bhat pointed out in his speech: "Real-time monitoring of voice is not easy, because it not only requires identifying the voice content, but also capturing the tone and intensity of the speaker, so as to accurately determine whether a passage is malicious." He further explained, "Context information is also a key factor in judging whether a discourse is offensive, which makes the problem more complicated."

　　To address these challenges, the company adopted a strategy of dividing illegal content into four main categories, and found that 85% of the illegal content belongs to these categories, and the "overwhelming majority" can be captured through 50 keywords. Bhat said: "If the top 50 keywords in these categories can be covered, the review work will be quite in place."

Heikinheimo, the head of voice security, emphasized the remarkable achievements made by AI supervision: after a year of operation, the system has covered 31 countries and regions, and the number of violations reported in each period of "active chat time" has been reduced by half. He mentioned that one of the notable advantages of AI is that it is emotionally unhealthy and does not exhaust, so machines are more efficient and stable than humans when dealing with "very obvious violations." However, he also admits that “humans are still stronger than machines” in understanding intentions and making judgments.

Heikinheimo explained: "In dealing with situations that may be close to the edge of judgment, or in rare cases where data is scarce and machine learning systems are difficult to train fully, the performance of human auditors is still better."