Study on Automatic Speech Recognition with the Hidden Markov Model Using the Genetic Algorithm

口语是人类与信息处理系统进行交流的最基本、最快捷、最方便的方式。语音自动识别是信息处理系统的语音感知功能。本研究的目的是开发一种基于遗传算法的高性能识别系统模型构建方法，并通过实验验证该方法的有效性。隐马尔可夫模型在语音自动识别中有着广泛的应用。然而隐马尔可夫模型有一个尚未解决的问题，即如何设计模型的最优结构。为了寻找HMM的最优结构，本研究提出将遗传算法作为自然进化过程的模型加以应用。在该算法中，随着生成的进行，性能较高的模型存活下来，可能性较低的模型死亡，最终得到全局最优结构。首先，我们将遗传算法应用于确定用于口语单词识别的离散HMM的结构。识别实验结果表明，无论是在封闭测试中还是在开放测试中，随着生成的深入，得到的结构都具有较高的识别分数。结果表明，该算法的识别分数高于目前最流行、性能最好的左-右结构，表明了遗传算法的有效性。然后，将遗传算法应用于连续HMM，得到与离散HMM相似的有效结果。作为该方法的改进版本，单词集编码方法、隐藏基因方法和某一状态下的交叉和突变方法都是有效的。

Spoken language is the most fundamental, fast and convenient method for human to communicate with information processing systems. An automatic speech recognition is the function of speech perception for the information processing system. The purpose of this research is to develop the model construction method for recognition systems with high performance using the genetic algorithm (GA), and to show the effectiveness of the method experimentally.The hidden Markov models (HMMs) are widely used for automatic speech recognition. However, the HMM has a problem still unresolved, i.e. how to design the optimal structure of the model.In order to search out the optimal structure of the HMM, we propose in this study the application of the GA which is the model of natural evolution process. In this algorithm, models with higher performance survive and models with lower likelihood die as the generation proceeds, then finally, the globally optimal structure is obtained.First, we applied the GA to the determination of the discrete HMM's structure for spoken word recognition. As a result of the recognition experiment, it was shown that the structures with higher recognition scores are obtained as the generation proceeds, not only in the case of closed tests but also open tests. The recognition score became higher than that of the Left-Right structure which is the most popular and with high performance, and the effectiveness of the GA was shown. Next, the GA was applied to the continuous HMM, and the effective result was obtained similarly to the discrete HMM. As the revised version of this method, the coding method of word set, the hidden gene method and the crossover and mutation in a state were shown to be effective.