人和机器的对话.pptVIP

车载语音控制和导航 针对车载噪声环境下的语音识别（控制命令/导航/电话） 嵌入式语音识别系统解码器的研制和开发 对噪声下连续语音的识别准确率达到90% 其它项目 * * * 人和机器的对话 我们研究的是什么？ 听 说 Sound, speech or music? sound speech music 言语链 由于语言的产生 …… 如果人类失去语言 …… 语言 交流的工具 每人每分钟说250字 每小时说话超过10000字 一张A4纸约1000字 中国3亿手机用户 = 海量数据 语音信息处理 多学科交叉 语音识别框架 数据——训练模型——测试新数据 Front-End Parameterization Cepstral parameters with smoothed first/second order derivatives BN Speaker Segmentation/Clustering divide audio into homogeneous segments Model Structure Hidden Markov Model using Gaussian Mixture distributions Left-to-right structure for speech HMM LVCSR Standard Phone Model Lexicon Design pronunciations per word common word variations Acoustic Models Context-dependent cross-word model (triphone/quinphone) Parameter tying (state-level/model-level) Decision tree-based state clustering Maximum likelihood estimation (MLE) for model parameter estimation LVCSR Language Models N-gram LM (e.g. Bigram/trigram/four-gram) Key issue is data sparsity Decoding Combine the LM probability with HMM likelihood Beamwidth searching Lattice structure (multiple tokens) nodes: word-ends in time arcs: word-word transitions LVCSR Advanced Tech. Discriminative Training Maximum mutual information estimation (MMIE) maximizes the sentence level posterior in log form Numerator is likelihood of data given correct transcription (as MLE) Denominator expands total likelihood in terms of all word sequences More closely related to word error rate than MLE Minimal word/phone error (MWE/MPE) Smoothed approximation to phone error in a word recognition context Minimize the Bayes’ Risk with particular loss function Advanced Tech. Speaker/Environment Adaptation Obtain the performance of a speaker/environ. dependent system with less data The mode can be incremental: the adaptation data is not all available in one block batch: the adaptation data is used in one block In addition for batch adaptation, the data my be supervised: the correct transcription is known (enrolment set) unsupe