SGER: Multi-Stream Approach Using Syllable Length Temporal Evidence in Acoustic Modeling of Conversational Speech

该项目研究使用调制频谱域中的流定义的自动语音识别(ASR)的多数据流方法。多流方法是一种新的声学建模技术，旨在提高ASR系统的鲁棒性。它从语音信号中提取不同的信息子流，然后在每个子流中进行独立的特征提取和概率估计。从每个子流获得的类别条件概率向量被合并以获得最终判决。利用子流中不同频谱区域的多频带模型已被证明对频率选择性退化具有较强的鲁棒性。该项目的目的是在不同的子流中利用不同范围的调制频谱。对不同调制光谱范围的相对重要性的研究表明，它们可能携带不同的信息。这使得调制频谱的各种元素成为多流方法的候选对象。这些特征的计算需要查看信号的较长片段(通常是音节长度片段)。该项目试图将特征向量元素的相对独立性和特征向量之间的中期时间相关性的新假设应用于ASR，以期提高其稳健性。

The project studies the multi-stream approach to automatic speech recognition (ASR) using stream definitions in the modulation spectrum domain. The multi-stream approach is a new acoustic modeling technique to increase the robustness of ASR systems. It derives different information sub-streams from the speech signal followed by independent feature extraction and probability estimation in each sub-stream. The vectors of class-conditional probabilities obtained from each sub-stream are merged to obtain the final decision. The multi-band model which utilizes different regions of the frequency spectrum in the sub-streams has been shown to be robust to frequency-selective degradation. The aim of the project is to utilize different ranges of the modulation spectrum in different sub-streams. A study of the relative importance of the different ranges of the modulation spectrum indicates that they might carry different information. This makes the various elements of the modulation spectrum interesting candidates for the multi-stream approach. The computation of these features requires looking at longer segments (typically syllable length segments) of the signal. The proposed project is an attempt to apply the new assumptions of relative independence of elements of the feature vector and medium-term time dependencies between feature vectors to ASR with a view to improve its robustness.