China Telecom Releases the First Large Speech Model Supporting 30 Dialects

Source: contribution
2024-05-26 10:32:52

China Telecom Artificial Intelligence Research Institute (TeleAI) recently released the first large-scale speech recognition model in the industry that supports the free mixing of 30 dialects - Xingchen Super Multi dialect Speech Recognition Large Model, which can simultaneously recognize and understand more than 30 dialects, such as Cantonese, Shanghai dialect, Sichuan dialect, Wenzhou dialect, etc. It is the largest speech recognition model supporting the most dialects in China at present.

It is reported that China Telecom Artificial Intelligence Research Institute has built a high-quality dialect database of more than 30 kinds and more than 300000 hours. The "distillation+expansion" joint training algorithm adopted by the R&D team can solve the problem of collapse of pre training under the conditions of large-scale multi scene data sets and large-scale parameters, and achieve stable training of 80 layer models of 1B parameters.

China Telecom said that the Star Speech Large Model is the first open source speech recognition large model based on discrete speech representation in the industry. Through the new modeling paradigm of "from voice to token to text", the speech transmission bit rate during reasoning is significantly reduced.

According to China Telecom, the Star Voice big model has been opened to the outside world and has been applied to Fujian, Jiangxi, Guangxi, Beijing, Inner Mongolia and other places of China Telecom Wanhao intelligent customer service pilot applications. By accessing the Star model, intelligent customer service can quickly understand 30 dialects, and handle about 2 million calls per day, greatly improving service efficiency and user experience.

Expand to read the full text
Click to join the discussion 🔥 (12) Post and join the discussion 🔥
This wonderful review
I think the rate is probably the recording of customer service calls, which is practical and not easy to cause disputes. Do not use the customer service recordings accumulated over the years, but monitor for training?
2024-05-27 09:22
two fabulous
report
That's because you think too much.
2024-05-27 09:22
one fabulous
report
twelve comment
two Collection
 Back to top
Top