tailieunhanh - Speech recognition using neural networks - Chapter 4

Related Research Bởi vì phát biểu công nhận về cơ bản là một vấn đề nhận dạng mẫu, và bởi vì các mạng thần kinh giỏi trong việc nhận dạng mẫu, nhiều nhà nghiên cứu đầu tự nhiên đã cố gắng áp dụng các mạng thần kinh để nhận dạng giọng nói. Những nỗ lực đầu tiên tham gia rất đơn giản hóa nhiệm vụ, ví dụ, phân loại các phân đoạn bài phát biểu như lồng tiếng / phát âm không được, hay mũi / fricative / plosive. Thành công trong những thí nghiệm này khuyến khích các. | 4. Related Research . Early Neural Network Approaches Because speech recognition is basically a pattern recognition problem and because neural networks are good at pattern recognition many early researchers naturally tried applying neural networks to speech recognition. The earliest attempts involved highly simplified tasks . classifying speech segments as voiced unvoiced or nasal fricative plosive. Success in these experiments encouraged researchers to move on to phoneme classification this task became a proving ground for neural networks as they quickly achieved world-class results. The same techniques also achieved some success at the level of word recognition although it became clear that there were scaling problems which will be discussed later. There are two basic approaches to speech classification using neural networks static and dynamic as illustrated in Figure . In static classification the neural network sees all of the input speech at once and makes a single decision. By contrast in dynamic classification the neural network sees only a small window of the speech and this window slides over the input speech while the network makes a series of local decisions which have to be integrated into a global decision at a later time. Static classification works well for phoneme recognition but it scales poorly to the level of words or sentences dynamic classification scales better. Either approach may make use of recurrent connections although recurrence is more often found in the dynamic approach. outputs Input speech pattern Figure Static and dynamic approaches to classification. Dynamic classification 51 52 4. Related Research In the following sections we will briefly review some representative experiments in phoneme and word classification using both static and dynamic approaches. . Phoneme Classification Phoneme classification can be performed with high accuracy by using either static or dynamic approaches. Here we review some typical .