Year: 2025
Type: article
Abstract: Automatic speech recognition (ASR) techniques are becoming more and more important in people’s daily life as a vital method of human-computer interaction. The ability to maintain a high recognition accuracy in noisy environments is the key for a model to be widely used in ASR. This article researches on the robustness against different kinds of background noise in Mandarin Chinese speech recognition and the baseline model used here is Bidirectional Long Short-Term Memory (BiLSTM). To compare the effects of different kinds of data augmentation method and different model structure, four common used data augmentation methods are applied in the process of training respectively and together, and two model methods, the CNNs and the attention mechanism, are combined with the baseline model, using the method of controlled experiment. After the model is trained, it will be tested within three kinds of background noise (car noise, café noise, white noise) to evaluate the anti-noise ability of different methods applied to the model. Among the four data augmentation methods, the method of random increased/decreased volume performed best in improving recognition accuracy, while the method of time-frequency masking increased Character Error Rate (CER) unexpectedly. As for the model methods, the CNNs performed better than the attention mechanism.
Author: Yatai Ji
Institution: Peking University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4409237086