数据==智能从语音识别技术谈起.PDF

数据==智能？ 从语音识别技术谈起 王东 清华大学校友人工智能论坛 2015-12-11 提纲 • 语音识别简史-从结构到数据 • 语音识别是否具有智能？ • 实现类人智能的一点思考 人类如何“听”声音  Sentences in brain (psychology)  Articulatory organs produce air pressure (physiology)  Speech wave transform in the air (transmission)  Auditory organs receive the wave (perception)  Brain decoding (neural science)  Understanding (linguistics, psychology…) 语音识别模拟人类听觉 / 主要困难：多样性和不确定性 Ellipses and Anaphors Limited vocabulary Multiple Interpretations Speaker Dependency Word variations VP NP NP Word confusability Context-dependency MY IS THREE ZERO SEVEN NUMBER SEVEN NINE TWO Coarticulation FOUR Noise/reverberation m I n m i s e th n I n z E o t s n ü e f O r r v r E Intra-speaker variability b r v n 5 如何解决不确定贯穿60年语音识别 进程 • 提取不变因素，去除噪声 – Feature pipe-line – Discriminative training – DNN • 用概率模型描述不确定性 – HMM,GMM • 复杂函数特定条件建模 – DNN,RNN