With the application of massive training data, speech recognition system has been able to recognize common speech more accurately Sibirch integrates supervised, semi supervised, self supervised and other methods, proposes a mixed supervision optimization scheme, fully exploits the value of data, continuously optimizes the speech recognition model, and obtains better speech recognition results The new automatic optimization system of "thousands of words and thousands of trainings" launched by Sibirch is an integrated automation solution that integrates the functional modules based on active learning, such as data screening, automatic bid submission, mixed supervisory acoustics and language model training, automated testing and online publishing In the future, Sibitch will continue to optimize the voice recognition link strategy, further shorten the update cycle of the general voice recognition model, meet the evolving business needs, and support more scene fields