Uyghur Acoustic Features Analysis And Implementation Of Command Word Recognition System

In recent decades, speech recognition technology had a rapid development,especially the command word recognition. Comparing with large vocabularycontinuous speech recognition, the task of the command word recognition is a simplerspeech recognition technology. It has been applied to many specific areas with moremature application. But there is few research achievements on this field, so we did tryin this respect, the main content of this paper includes:1. We use statistical and experimental phonetics knowledge to analyze eachphoneme’s formant frequency, duration and intensity and classified them by speechrate.2. Fuse formant in the original features, modify the number of state based onduration, and then analyze the confusion matrix for checking the recognition error.3. We estimate the acoustic models’ initial parameters with a small amount ofmanually labeled continuous speech corpus, and train acoustic models by a largenumber of non-manual corpus. We use the maximum likelihood estimation (MLE)training acoustic model. Finally, we build state-tied HMM triphone models based ondecision tree.4. We build grammar based on rules for guiding command word recognition.5. We add the process module of Uyghur to Aitalk4, and use filler model incommand word recognition.6. The system contains six main functions. We can use it for calling, messagesending, weather forecasting getting, and file opening. The final test recognition resultis81.4%.