基于分形维的连续语音实时分割技术

为能准确有效地判断出连续语音中各个音节的起止点,提出了一种改进的分形维计算方法,该算法将插值分形维的步长因子进一步精确到采样频率的数量级上,先求出语音分形维的最小二乘能量轨迹,再差分求其动态特征;在此基础上,设计了连续语音的两级搜索实时分割算法,并进行了基于DSP的硬件系统实验。结果证明,该算法较好地实现了语音段的实时分割和汉语音节切分,鲁棒性好,使得系统在信噪比为0 dB时音节分割准确率仍可保持在一个较高的水平上。最后开发了一个在线汉语语音标注器,借此阐述了论文工作在语音识别方面的应用。

Continuous Speech Real-time Segmentation Technology Based on Short Time Fractal Dimension

An improved algorithm for the calculation of fractal dimension is proposed in order to efficiently detect the endpoint of syllable,which gives full consideration to the characteristics of speech signals in time domain and frequency domain.First,least square''s energy trajectory of speech is got,and then its difference is made to obtain dynamic characteristics.On this basis,two-level searching algorithm for real-time segmentation of continuous speech is designed and experiments are made based on DSP system.The results show that this method is time-saving and efficient,which shows a better result for the real-time segmentation of speech segment and the division of Chinese syllable,even in a low signal-to-noise ratio(0dB).Finally,an online Chinese-mark system is developed to illustrate the application of the research in speech recognition.