论文部分内容阅读
传统的PRSVM方法存在以下问题:音素识别器的符号化结果与原语音存在不一致;向量空间维数高,稀疏.针对以上问题,先改用更适合噪声环境下连续电话语音的音素识别器,并采用词图结构改善解码效果,再分别用全局和局部两种隐含语义分析策略改进区分性训练问题.实验表明,本方法不但有效,而且大大减少了运算量.在NIST2007语种识别评测30秒、10秒和3秒任务中,本方法比基线系统性能有显著提高,等错误率分别相对降低了22.3%、14.7%和12.2%.
The traditional PRSVM method has the following problems: the symbolic result of the phoneme recognizer is inconsistent with that of the original voice; the dimension of the vector space is high and sparse. In view of the above problem, the phoneme recognizer, which is more suitable for the continuous phone voice in noisy environments, is used instead The word graph structure is used to improve the decoding effect, and then the global and local implicit semantic analysis strategies are respectively used to improve the discriminative training.Experiments show that the proposed method is not only effective, but also greatly reduces the computational complexity.In the NIST2007 language recognition evaluation of 30 seconds, In the 10s and 3s tasks, the performance of this method has been significantly improved compared to the baseline system, with equivalent error rates of 22.3%, 14.7% and 12.2%, respectively.