Recently, Huya Live’s first live broadcast real-time silencer system has been launched and has been applied on its platform. The system provides comprehensive supervision capabilities for the “pre-prevention and control” in live broadcast scenarios.
It is reported that the current mainstream content security review method is AI + manual review, which is a "post-event review" processing method.
The system developed by Huya is a technology that is first reviewed and released. Based on Huya's self-developed audio algorithm and multimedia processing platform leaf, it can perform real-time monitoring of illegal audio during live broadcast without increasing the delay of live broadcast. Silencing, effectively reducing or even completely blocking the spread of risky content, realizing real-time shielding, reviewing before publishing.
"For the scenario application of AI capabilities in content risk control, we give priority to the application of live audio scenarios with high manual review difficulty and slow efficiency. In the next step, we will try it in live video scenarios." Huya Risk Control said the team leader.
In order to achieve no delay in the live broadcast scene, the Huya Dopamine AI technology team has made a lot of optimizations on the speech recognition model and decoding module, and the decoding of each speech segment can achieve a stable and consistent recognition time.
"This is very important, because the large fluctuation of the decoding time of audio clips will lead to the leakage of illegal audio. On a common 2.1G main frequency CPU, our real-time rate reaches 0.08, which is equivalent to only 80ms for 1s audio to be recognized.", The person in charge of Huya AI noise reduction technology said.
The complexity of the live broadcast scene is larger than that of the general speech recognition scene, and the recognition accuracy of the complex scene has always been a difficulty in the industry.
"Low accuracy will cause large-scale false silencing of live broadcasts, reducing user experience, and low recall rates will lead to leakage of illegal voices. In order to achieve the goal of high recall and high accuracy, the Huya Dopamine team has developed a VAD based on live broadcast scenarios. Algorithms, speech recognition algorithms, post-processing algorithms, and a large number of samples of complex scenes are collected at the same time, and the algorithm is iteratively optimized, so that we have a high recognition accuracy and recall rate in complex scenes such as live broadcasts. On the other hand, AI The optimization of the model still relies on sample calibration work, which is extensive, systematic, long-term and meticulous." said the above-mentioned person in charge.
When the Cyberspace Administration of China deployed the "Qinglang" series of special actions in 2022, it pointed out that the action focused on 10 aspects, including live webcasting, information content chaos, online rumors, and the online environment for minors.
For every Internet content platform, ensuring the legitimacy and standardization of platform content and improving the platform content review mechanism have become issues that must be paid attention to.
According to the data, the "Sky Eye" AI engine developed by Huya in 2015 combines cutting-edge technological achievements such as artificial intelligence and computer vision with the security of Internet content.
The system can empower AI capabilities for different scenarios, realize the landing innovation of intelligent recognition (including audio and video, images, text) and business risk control, make traditional content security work more efficient and cost-effective, and realize automatic risk prediction .
In addition, the "Huya Tianyan Content Security SaaS Solution" has been launched on Amazon Cloud, enabling the online audio-visual industry.
The person in charge of the content risk control of Huya live broadcast said that the platform has been promoting the content security work in depth, adopting the trinity model of "AI intelligent identification, manual review and network volunteers".
The Tianyan real-time silencer system is an expansion of the platform's AI security application. The platform will build a comprehensive security attack and defense system and a more accurate content review system, providing reference samples and cutting-edge solutions for the construction of a healthy web live content ecosystem.