汉语元音的非线性预测

最近研究表明非线性的局部模型是最精确的混沌时间序列预测方法之一．同时有研究表明汉语的产生过程包含非线性．传统的语音处理技术忽视了语音中可能存在的非线性结构，因而限制了处理能力的进一步提高．对汉语元音／a／和／u／的细致相空间分析结果表明，汉语元音的重构相空间有类似于混沌吸引子的结构．根据汉语元音发声过程的非线性特性，进一步对其进行相空间的局部线性预测分析，计算了预测误差随预测步长的关系曲线，讨论了嵌入维数对预测性能的影响，并且与传统的线性方法作了比较．结果表明，尽管非线性的局域线性预测法存在计算开销大的问题，但其预测视野要远大于线性预测法，总体预测性能明显优于线性预测法．

Nonlinear Prediction of Chinese Vowels

Source-filter models are traditional models for production of speech. Usually, the filter is linear and based on linear prediction; the excitation is either modeled as white noise, or a simple pulse train. While this approach has led to great advances in the last 30 years, it neglects nonlinear structure known to be present in the speech signals. In practical applications, this manifests itself as an increase in bit rate, less natural speech synthesis and an inferior ability to discriminate speech sounds. Previous studies have shown that local models are among the most accurate methods for predicting chaotic time series. Recently, it has also been found out that nonlinearities exist in both vowels and consonants of Chinese speech. In this work, the locally linear model has been introduced to solve the problem of modeling and prediction of Chinese speech. The state space structure has been investigated in detail for vowels /a/ and /u/, the result of which shows that the reconstructed state space structure of vowels has the same properties as those of chaotic time series. By making use of the feature of low-dimensional nonlinearity in the sounding process of Chinese vowels, the locally linear prediction method originally developed from nonlinear dynamics is applied to the prediction of Chinese vowels. Some important aspects such as the prediction error vs. time, the influence of the embedding dimension, and the prediction horizon, are investigated and analyzed in detail. The nonlinear predictability is compared with that of the traditional linear method, namely, the iterated linear prediction. The result shows that the prediction horizon of the nonlinear method is farther than that of the linear one, and moreover, the overall performance within the prediction horizon excels that of the linear prediction, though the computation of the former is somewhat heavier than that of the linear method.