轨道交通车载端到端语音合成

An end-to-end text-to-speech system for vehicle-mounted devices

高自然度的语音合成是车载人机交互进入高级智能化的基本要求之一。现阶段的轨道交通领域仍在广泛使用传统的低自然度语音合成算法,这与高速发展的智能化人机交互技术脱节。相比之下,端到端的深度学习语音合成算法凭借其近乎媲美人声的自然度已经成为各领域语音合成的主流算法。文章介绍了一种适用于离线轨道交通车载环境的端到端深度学习语音合成算法,该算法的主观意见评分达到4.18,并且在车载嵌入式硬件平台英伟达Xavier上的实时率达到0.52。试验证明,该算法不仅在自然度等主观性能上远超传统语音合成算法,同时也具备在轨道交通离线车载环境下的工程实用性。

High-naturalness text-to-speech is one of the basic requirements for advanced intelligence in vehicle-mounted humanmachine interaction.Currently,in the rail transit field,there is widespread use of traditional low-naturalness text-to-speech algorithms,which are out of touch with the rapidly developing intelligent human-machine interaction technology.In contrast,end-to-end deep learning-based text-to-speech algorithms,with their nearly human-like naturalness,have become dominant in various fields of text-to-speech.This paper introduced an end-to-end deep learning-based text-to-speech algorithm suitable for offline railway vehicle environments.The mean opinion score of this algorithm reached 4.18,and the real-time rate on the vehicle-mounted embedded hardware platform NVIDIA Xavier reached 0.52.Experiments show that this algorithm not only outperforms traditional text-to-speech algorithms in terms of subjective performance such as naturalness,but also possesses engineering practicality in the offline vehicle environment of railway transportation.