| Summary: | 碩士 === 國立中興大學 === 統計學研究所 === 98 === This paper is to discuss the speech recognition of 337 isolated mandarin words for speaker-dependent and use the method of shifting frames to recognize. First, I record the 337 isolated mandarin words ten times, and save them to speech database. After recording, I focus the speech database on pre-processing, and then through the linear prediction coding, the cepstrum coding to obtain the speech features.
With the speech features, I use the weight (0,1) to find out the vowel. And then, I use the vowel with the optimum shifting frames of the K-nearest neighbor to find the test will belong to which isolated mandarin words.
In this experiment, it will be divided into three types of shifting frames to recognize the 337 mandarin words in the vowel, move 0 units, move 1 units, move 2 units, respectively. The best recognition rate of three types of shifting frame is to move 0 units. The recognition rate of Top1 is 98.02% and the highest overall recognition rate is 89.81%.
|
|---|
