端到端的藏语语音合成方法

近年来，得益于计算机运算能力的提高和语音数据的不断积累，涌现出许多基于机器学习的语音处理新技术，其中基于深度神经网络算法，端到端的Tacotron2语音合成系统框架得到业界广泛的青睐。它是一个开源程序，简单易行，已成功地应用于多种语言和不同音色的语音合成。该文研究Tacotron2在藏语中的应用，取得了良好的实验结果。首先，通过自然语音采集、自动标注、声学分析等构建了一个中等规模(5500句)藏语卫藏方言的语音语料库，其中包括藏文音素转写、特殊符号处理和Mel谱等各项数据；其次，利用开源程序Tacotron2和上述语音库进行了藏语语音合成试验；最后，通过对合成语音和自然语音的偏差分析，和对合成语音的自然度的主观评价，表明了基于端到端的藏语语音合成方法有效地减少合成语音的频谱蜕变，提升了合成语音的自然度。因此，基于“端到端”的Tacotron2合成框架在藏语语音合成中具有重要的应用价值，值得进一步研究和推广应用。     

Automatic measurement device of laser-induced damage threshold and its experiment北大核心

A set of 1 064 nm and 532 nm automatic measurement device of laser-induced damage threshold (LIDT) of dual-wavelength optical elements was designed and constructed for automatic detection of LIDT of optical elements film. The device was mainly composed of pulsed laser light source, beam parameter diagnosis component, damage online diagnosis component, scanning motion platform for parts to be tested and control system. The whole measuring device and measurement process were automatically controlled by the computer integrated measuring software based on Labview, which could realize the automatic measurement of damage threshold in the range of energy density from 0.1 J/cm2～100 J/cm2, and the device was also used to measure the 1 064 nm antireflection film and aluminum reflection film samples. The results show that the damage thresholds are 27.09 J/cm2 and 3.21 J/cm2, with a relative uncertainty of 3.91% and 5.61%, respectively. © 2023 Editorial office of Journal of Applied Optics. All rights reserved.