Robust HMM speech recognition using robust time-varying complex speech analysis

我们已经提出了几种鲁棒时变复杂AR（TV-CAR）语音分析方法，我们打算通过将TV-CAR方法作为语音识别的前端来实现鲁棒语音识别。TV-CAR方法采用时变复AR模型作为语音产生模型，其中AR参数用复基展开表示。TV-CAR方法可以估计分析语音信号的时变复AR参数。在2002年之前，我们已经提出了MMSE、m估计、IV、GLS（一般最小二乘）和ELS（扩展最小二乘）方法。GLS和ELS方法可以估计无偏和噪声影响较小的语音频谱，实现鲁棒性语音频谱估计。自2002年以来，我们提出了更精确的语音分析，基于前向和后向线性预测（FB-LP）的GLS和ELS算法以及基于输出误差的ELS算法。我们采用HTK（HMM Tool Kit）作为HMM语音识别。为了将TV-CAR方法应用到HTK中，我们研究了TV-CAR参数到HTK格式的LPC倒谱系数（LPCC）的参数转换，从而实现了用TV-CAR方法进行HTK语音识别。目前，我们正在评估时变特征和复杂分析在HTK语音识别中的有效性。此外，我们将评估鲁棒语音分析算法的有效性，即基于ELS和基于FBLP的ELS。

We have already proposed several robust time-varying complex AR(TV-CAR) speech analysis methods and we intend to realize robust speech recognition by means of adopting the TV-CAR method as a front-end of speech recognition. The TV-CAR methods adopt time-varying complex AR model as a speech production model in which AR parameter is represented by a complex basis expansion. The TV-CAR methods can estimate time-varying complex AR parameters for analytic speech signal. We have already proposed MMSE, M-estimation, IV, GLS(General Least Square) and ELS(Extended Least Square) method before 2002. A GLS and ELS method can estimate unbiased and less noise effected speech spectrum and can realize robust speech spectrum estimation. Since 2002, we have proposed more precise speech analysis, forward and backward linear prediction(FB-LP) based GLS and ELS algorithms and output error based ELS algorithm. We adopt HTK(HMM Tool Kit) as HMM speech recognition. In order to apply the TV-CAR method to the HTK, we have investigated parameter conversion from TV-CAR parameters to the HTK formatted LPC cepstrum coefficients(LPCC), as a result, we have realized HTK speech recognition using the TV-CAR method. Now we are evaluating the effectiveness of time-varying feature as well as complex analysis on HTK speech recognition. Furthermore, we will evaluate the effectiveness of robust speech analysis algorithm, viz. the ELS and FBLP based ELS.