Power Speech Feature Output Model Based on Convolution Algorithm
Konferenz: ECITech 2022 - The 2022 International Conference on Electrical, Control and Information Technology
25.03.2022 - 27.03.2022 in Kunming, China
Tagungsband: ECITech 2022
Seiten: 4Sprache: EnglischTyp: PDF
Autoren:
Chen, Jinlong; Yang, Shuai; Li, Wudi; Guo, Meng; Xu, Sheng; Li, Yujing (Power Dispatch Control Center, Guizhou Power Grid Co., Ltd., Guizhou, Guiyang, China)
Inhalt:
In the development of power industry, power data communication network plays a key role in power grid management informatization, dispatching intellectualization and operation marketization, and it is an important part of power information infrastructure. The voice communication feature analysis in the process of transmitting information can maintain strong timeliness and transmit more complex data types. This paper uses a speech recognition and speech synthesis module to complete the application of voice interaction technology under the distribution station platform. Transfer learning is introduced to improve the recognition rate of a small number of words in the power distribution station. At the same time, a Transformer model based on feature words is built in the acoustic model, which improves the transcription ability and speed of the acoustic model from Pinyin to text. Experiments show that this method has great technical advantages in accelerating the decoding process from spectrum to waveform. Compared with other methods, the training speed and inference speed are increased by 3.4% and 5.6% respectively, and the vocoder generation effect is better than that of the comparison method, which improves the recognition accuracy. Using voice to control the distribution station can greatly improve the intelligence of the distribution station, and it is also more convenient and fast. Therefore, the application of this method in the smart substation has important economic value.