Robust Syllable Recognition in the Acousic-Waveform Domain

该建议涉及声学波形领域中的语音单元(音素和辅音-元音音节)的稳健分类/识别。这项研究的动机来自于这样的想法，即语音单元在由声波波形形成的高维空间中应该比在较小的表示空间中分离得更好，该空间在最先进的语音识别系统中使用，并且涉及显著的压缩和降维。因此，与低维特征空间中的分类相比，声波形域中的识别/分类应该表现出更高的对加性噪声的稳健性。在项目的第一阶段，我们将研究在严重噪声条件下，在声波形域中的语音单元分类，大约在0dB信噪比及以下，而在第二阶段，我们将研究使分类对线性滤波也具有稳健性的技术。第一阶段要解决的具体任务可以概括如下：1.研究单个语音单元的声学波形集合的详细结构；特别是它们的内在维度，以及数据集中在其上的可能的非线性表面的存在。以上述第1项的研究结果为指导，估计声学波形域中语音单元分布的统计模型。然后，我们将设计和系统地评估所谓的生成分类器，其定义属性是它们基于这样的统计模型。使用判别性分类技术(人工神经网络、支持向量机和相关向量机)研究声学波形域中语音单元的分类。这些方法可以替代生成技术，因为它们直接集中在分类问题上，而不需要为每个语音单元建立显式的波形分布模型。通过对语音单元进行分层分组来构建分类器。将构建顶级分类器来区分一小部分相似语音单元的组，然后是将组分为子组的分类器，依此类推。我们将探索不同的方法来定义子组，包括项3中分类器的混淆矩阵，项2中获得的统计模型之间的适当距离度量，以及可能的感知实验。反对我们方法的一个潜在论点是，在存在线性滤波的情况下，声波波形域的分类将崩溃。然而，这可以通过考虑窄带信号来避免：对于这些信号，线性滤波的效果大致相当于幅度缩放和时间延迟。因此，在项目的第二阶段，我们将考虑使用声学波形的窄带分量进行语音分类。对于个别子带中的信号分类，将考虑在项目第一阶段研究的技术。一个新的问题是如何合并子带分类器的结果以最小化总体分类误差。这里将使用最近开发的机器学习技术，如支持情况中所规定的。如所解释的，单个子带分类器应该对线性滤波具有健壮性，因为后者不会显著改变窄带信号的形状。另一方面，子带波形的空间维度仍然足够高，以便于分类对加性噪声具有健壮性。因此，整个方案预计对加性噪声和线性拟合都是稳健的。

This proposal is concerned with robust classification/recognition of speech units (phonemes and consonant-vowel syllables) in the domain of acoustic waveforms. The motivation for this research comes from the idea that speech units should be much better separated in the high-dimensional spaces formed by acoustic waveforms than in the smaller representation spaces which are used in state-of-the-art speech recognition systems and which involve significant compression and dimension reduction. Hence, recognition/classification in the acoustic waveform domain should exhibit a higher level of robustness to additive noise than classification in low-dimensional feature spaces.In the first phase of the project we will investigate classification of speech units in the acoustic waveform domain under severe noise conditions, around 0dB signal-to-noise ratio and below, while in the second phase we will study techniques which would make classification robust also to linear filtering. The particular tasks that will be tackled in the first phase can be summarized as follows:1. Study the detailed structure of the sets of acoustic waveforms of individual speech units; in particular their intrinsic dimensions, and the existence of possible nonlinear surfaces on which the data are concentrated.2. Guided by the findings from item 1 above, estimate statistical models of the distribution of speech units in the acoustic waveform domain. We will then design and systematically assess so-called generative classifiers, whose defining property is that they are based on such statistical models.3. Investigate classification of speech units in the acoustic waveform domain using discriminative classification techniques (artificial neural networks, support vector machines, and relevance vector machines). These can be a useful alternative to generative techniques because they focus directly on the classification problem without building explicit models of waveform distributions for each speech unit.4. Construct classifiers by grouping speech units hierarchically. Top-level classifiers will be constructed to distinguish between a small of groups of similar speech units, followed by classifiers separating groups into subgroups and so on. Different methods for defining subgroups will be explored, including confusion matrices of the classifiers from item 3, appropriate distance measures between the statistical models obtained in item 2, and possibly perceptual experiments.A potential argument against our approach is that classification in the acoustic waveform domain will break down in the presence of linear filtering. However, this can be avoided by considering narrow-band signals: for these, the effect of linear filtering is approximately equivalent to amplitude scaling and time delay. In the second phase of the project, we will therefore consider speech classification using narrow-band components of acoustic waveforms. For classification of signals in individual sub-bands, the techniques investigated in the first phase of the project will be considered. A new issue is then how to combine the results of sub-band classifiers to minimize the overall classification error. Here recently developed machine learning techniques will be used, as specified in the case for support.As explained, individual sub-band classifiers should be robust to linear filtering because the latter does not significantly alter the shape of narrow-band signals. On the other hand, the dimension of the spaces of sub-band waveforms will be still high enough to facilitate classification robust to additive noise. Hence, the overall scheme is expected to be robust to both additive noise and linear fitering.