Aiming at the problem that the recognition accuracy of the system will decline when the target speaker voice tested in the speech recognition system and the speaker voice of the training data are significantly different, a power system speech recognition method based on natural language processing is proposed. It is speaker adaptation in the feature space. By adding speaker identity vector I-Vector auxiliary information to the DNN acoustic model, the speaker difference information in the feature is removed, the impact of speaker difference is reduced, and semantic information is retained. The experimental results on TEDLIUM open source data set show that the word error rate WER of the system is 7.7% and 6.7% higher than that of the baseline DNN acoustic model when the feature of this method is fbank and fMLLR respectively.
Power System Speech Recognition Method Based on Natural Language Processing
2023-10-11
2522571 byte
Conference paper
Electronic Resource
English
Language processing for Chinese speech recognition
IEEE | 1994
|Language Processing for Chinese Speech Recognition
British Library Conference Proceedings | 1994
|SPEECH RECOGNITION DEVICE, SPEECH RECOGNITION SYSTEM, AND SPEECH RECOGNITION METHOD
European Patent Office | 2019
|SPEECH RECOGNITION DEVICE, SPEECH RECOGNITION SYSTEM, AND SPEECH RECOGNITION METHOD
European Patent Office | 2020
|SPEECH RECOGNITION DEVICE, SPEECH RECOGNITION SYSTEM, AND SPEECH RECOGNITION METHOD
European Patent Office | 2020
|