基于Conformer的端到端中英文管制语音识别

End-to-End Chinese-English Speech Recognition Based on Conformer in Air Traffic Control

将语音识别技术应用到空中交通管理系统中可以提高飞行安全并降低管制员的工作负荷,目前已有的管制语音识别技术在中英文识别上效果较差,因此提出了一种基于Conformer-CTC/Attention的中英文管制语音识别框架。该方法使用基于改进的Conformer共享编码器对输入序列进行语言分类并以参数有效的方式对音频序列的局部和全局相依性进行建模,添加了语种分类模块来判断输入语音序列的语种,还采用了CTC解码器和注意力解码器联合解码的多任务建模方法。最后在建立的民航数据集对所提出的框架进行验证,试验结果表明,Conformer-CTC/Attention(Language-Category)相对于基线模型错误率降低,识别效果达到预期。

Applying speech recognition technology to air traffic management systems can enhance flight safety and reduce the workload of air traffic controllers.Currently,existing speech recognition technologies for air traffic control perform poorly in recognizing both Chinese and English.Therefore,a Conformer-CTC/Attention-based framework for Chinese-English air traffic control speech recognition is proposed.This method utilizes aimproved Conformer-based shared encoder for language classification of input sequences and effectively models local and global dependencies in audio sequences.It incorporates a language classification module to determine the language of the input speech sequence.It also employs a multi-task modeling approach with both CTC and attention decoders for joint decoding.Finally,the proposed framework is validated on a constructed civil aviation dataset.Experimental results indicate that Conformer-CTC/Attention(Language-Category)achieves a lower error rate compared to the baseline model,demonstrating the expected recognition performance improvement.