| Speech is the most widely used and the most significant information carrier in mutual communication. In practical environment, speech is inevitably interfered by noise. The noise not only degrades the quality and the intelligibility of speech, but also worsens the capability of the system. With the increase of practical application and the further study of speech processing technology, it is necessary for various speech processing systems to improve their capability. As speech enhancement technology is widely used in enhancing the anti-noise capability, the input Signal-to-Noise Ratio of the speech processing system and improving the quality and intelligibility of speech, speech enhancement technology is usually used as the pre-processing module in the speech processing system. Consequently, it is significant to study speech enhancement technology.In this thesis, we firstly illustrate the speech enhancement algorithm and its present research situation, and then point out that the spectral subtraction method and its various improved algorithms, which are based on the short-time spectrum estimate, are commonly used as the speech enhancement algorithms because of their less calculation complexity and very effective enhancement effect, therefore we decide to employ the spectral subtraction method as the basic algorithm of the real-time speech enhancement system. Therefore, we labor and explicate the spectral subtraction method and its six improved algorithms based on the analysis of the characteristics of speech and noise and human ear sensation characteristic. Finally, in view of existing research problem of the speech enhancement, we employ the last improved algorithm of the spectral subtraction method Martin algorithm, which is the spectral subtraction method based on the Kalman filtering, as the processing algorithm of the real-time speech enhancement system. The Martin algorithm, which is the spectral subtraction method based on the Kalman filtering using the double AR model of speech and noise on the assumption that the speech is contaminated by colored noise, can suppress and eliminate noise effectively. Furthermore, we use C programming language and MATLAB software to get the computer simulating result, in which we use the Martin algorithm to enhance the noisy speech with SNR of-5dB, 0dB, +5dB, +10dB respectively and test their enhancement effect. Judging from the computer simulating result, we can clearly know that the Martin algorithm is very efficient in enhancing the SNR and improving the intelligibility of speech.On the basis of the above computer simulating result, we use C programming language and assembler language on the hardware platform of TMS320VC5416 DSP and TLV320AIC23 Codec to realize many technologies such as sampling in the form of frame, the move of the frames, the overlap of seventy five percent among frames, the synchronous working condition among CPU and double DMA which are working under the ABU model, the interrupt service routine calling and the multiplying program block of FFT and IFFT and so on. Moreover, we built the real-time realization platform system based on these technologies and then real time enhance the noisy speech by Martin algorithm using the transformation mechanism between integer number and decimal number on the real-time realization platform system.In the end, the technical parameters of the real-time realization platform system are provided, as well as the SNR of the output enhanced speech when the input noisy speech with corresponding SNR of-5dB, 0 dB, +5 dB. The increase of the SNR by a large margin compare output enhanced speech with the input noisy speech and the audition experiment of the real-time speech enhancement processing system indicate that the real-time system can enhance the SNR and the MOS value, and improve the intelligibility and quality of speech effectively. Consequently, this real-time speech enhancement system can greatly satisfy the actual demand of the real-time application. |
PDF Full Text Request