基于Ca-GAN增强的机坪管制指令识别方法研究

Research on Apron Control Command Recognition Method Based on Ca-GAN

摘要:

我国枢纽机场长期处于繁忙状态，高负荷带来信息交互失真的风险，语音识别技术可用于辅助决策，然而管制语音特殊性及样本量局限性使传统深度学习技术难以直接应用于机坪管制领域。针对这一问题，提出了一种基于小样本学习的语音识别方法。首先提出数据增强方法，通过结合先验领域知识，构建基于数据生成策略组的生成对抗网络来增强声学模型识别能力来进一步提升模型效果；然后通过重构声学模型部分结构和参数；最后通过迁移学习方法将通用语音库中的声学建模特征应用到机坪管制语音指令的识别中。实验结果表明，该方法将字错率减少至6.14%。该研究可应用于机场高级地面活动引导及控制系统中机坪管制语音指令的检测和识别，助力现代机场高质量运行。

Abstract:

China's hub airports have been busy for a long time, and the high load brings the risk of distorted information interaction. Speech recognition technology can be used to assist decision-making, however, the special characteristics of control speech and sample size limitations make it difficult to apply traditional deep learning techniques directly to the ramp control field. To address this problem, a speech recognition method based on small sample learning is proposed. Firstly, a data enhancement method is proposed to further improve the model effect by combining a priori domain knowledge and constructing a generative adversarial network based on a data generation strategy group to enhance the acoustic model recognition ability; then, some structures and parameters of the acoustic model are reconstructed; finally, acoustic modeling features from a general speech library are applied to the recognition of ramp control speech commands by a migration learning method. The experimental results show that the method reduces the word error rate to 6.14%. This research can be applied to the detection and recognition of ramp control speech commands in advanced ground activity guidance and control systems in airports, which can help modern airports operate with high quality.