Speech Recognition for Power Dispatching Scenarios

In power dispatch speech recognition, the recognition scenario is different from quiet public domains, as it involves challenges such as noisy audio signals and an abundance of technical terms. Therefore, this paper proposes an acoustic model framework to enhance the attention mechanism, using the Conformer as the sound encoder and introducing the Deep Residual Shrinkage Network to process audio signals. In the decoder part, the decoding results of CTC are fed into the Attention network for scoring again, obtaining new score rankings to achieve the output with the highest probability. Finally, a shallow fusion acoustic model based on statistical language models n-gram is utilized to further improve the overall recognition capabilities of the model for power dispatch scenario audio. A comparison with baseline models on public Mandarin datasets and power dispatch scenario audio datasets showed a reduction in Character Error Rate (CER) by 2.3 and 4.1 respectively, validating the advanced nature of the model.