Abstract: In order to make the speech air conditioner not only support the recognition ability of a specific dialect, but also accurately and intelligently judge the dialect type used by users, the Attention-based Encoder-Decoder end-to-end system algorithm is preferred in the overall model architecture. Combined with SMAD technology, the end-to-end recognition system of cross dialect speech is built by using micro representation methods such as speech pronunciation characteristics and pronunciation positions, and the joint training method of dialect and wake-up words based on the conformer-transformer, which finally realizes the modeling of dialect recognition engine based on small data set, and realizes the functions of dialect wake-up and language adaptive judgment on the small resource system.

Key words: Dialect recognition, AED system, SMAD technology, Cross dialect speech recognition, Joint training methods, Air conditioning applications