China Telecom Releases Industry-Leading Large-Scale Dialect Speech Recognition Model, Supporting 30 Dialects Mixed Speech

2024-05-27

Recently, TeleAI, the artificial intelligence research institute under China Telecom, has made a significant breakthrough in the field of speech recognition technology. They have successfully released the industry's first large-scale speech recognition model, the Starry Sky Multi-Dialect Speech Recognition Model, which supports the free mixing of 30 dialects. This innovative achievement not only breaks the limitation of single-model dialect recognition but also provides strong technical support for language services nationwide.


It is reported that the Starry Sky Multi-Dialect Speech Recognition Model can simultaneously recognize and understand more than 30 dialects, including Cantonese, Shanghainese, Sichuanese, and Wenzhounese, making it the speech recognition model that supports the most dialects in China. This model is undoubtedly a blessing for elderly people and users in remote areas, greatly improving their convenience and efficiency in accessing information services.

In terms of technical research and development, the Starry Sky Speech Model development team has adopted the "distillation + expansion" joint training algorithm, successfully solving the problem of pre-training collapse under the conditions of super-large-scale multi-scenario datasets and large-scale parameters. They have achieved stable training of a 1 billion parameter 80-layer model. This technological innovation injects new vitality into the development of speech recognition technology.

Currently, the Starry Sky Speech Model has been widely used in China Telecom's Wanhao Intelligent Customer Service and has been piloted in Fujian, Jiangxi, Guangxi, Beijing, Inner Mongolia, and other regions. This intelligent customer service can "understand" 30 dialects in seconds and can handle approximately 2 million calls per day, greatly improving the efficiency and quality of customer service.

China Telecom's Artificial Intelligence Research Institute plans to further expand the support range of the Starry Sky Speech Model and build the first large-scale speech recognition model that covers 333 dialects and major minority languages across the country. The achievement of this goal will further promote China Telecom's innovation and development in the fields of intelligent customer service, smart governance, and smart home.

According to previous reports, China Telecom had 413 million mobile users in April, with an addition of 1.8 million users. The number of 5G package users reached a staggering 332 million, with an addition of 2.9 million users. At the same time, the number of wired broadband users and fixed-line telephone users also maintained steady growth. These data not only reflect China Telecom's leading position in the field of communication services but also provide a solid user base for its development in emerging fields such as artificial intelligence.