基于声学特征空间非线性流形结构的语音识别声学模型

从语音信号声学特征空间的非线性流形结构特点出发, 利用流形上的压缩感知原理, 构建新的语音识别声学模型. 将特征空间划分为多个局部区域, 对每个局部区域用一个低维的因子分析模型进行近似, 从而得到混合因子分析模型. 将上下文相关状态的观测矢量限定在该非线性低维流形结构上, 推导得到其观测概率模型. 最终, 每个状态由一个服从稀疏约束的权重矢量和若干个服从标准正态分布的低维局部因子矢量所决定. 文中给出了局部区域潜在维数的确定准则及模型参数的迭代估计算法. 基于RM语料库的连续语音识别实验表明, 相比于传统的高斯混合模型(Gaussian mixture model, GMM)和子空间高斯混合模型(Subspace Gaussian mixture model, SGMM), 新声学模型在测试集上的平均词错误率(Word error rate, WER)分别相对下降了33.1%和9.2%.

Feature Space Nonlinear Manifold Based Acoustic Model for Speech Recognition

Based on nonlinear manifold structure of acoustic feature space of speech signal, a new type of acoustic model for speech recognition is developed using compressive sensing. The feature space is divided into multiple local areas, with each area approximated by a low dimensional factor analysis model, so that in a mixture of factor analyzers is obtained. By restricting the observation vectors to be located on that nonlinear manifold, the probabilistic model of each context dependent state can be derived. Each state is determined by a sparse weight vector and several low-dimensional factors which follow standard Gaussian distributions. The principle for selection of the dimension for each local area is given, and iterated estimation methods for various model parameters are presented. Continuous speech recognition experiments on the RM corpus show that compared with the conventional Gaussian mixture model (GMM) and the subspace Gaussian mixture model (SGMM), the new acoustic model reduces the word error rate (WER) by 33.1% and 9.2% respectively.