MFCC Feature Extraction Research Based On ICA And Its Implementation On DSP

The speech recognition systems perform excellently in the laboratoryenvironment, while work poorly in the actual noisy environment, with the recognitionrate and robustness declining sharply. Speech feature extraction is essential in speechrecognition research, and affects the robustness of the speech recognition systemssignificantly; on the other hand, with the wide application of mobile portable devices,it is urgent that these embedded devices be configurated with efficient speechrecognition function.In order to improve the robustness of the embedded speech recognition systems,the thesis adopts a data-driven feature transformation method-IndependentComponent Analysis (ICA) to improve the commonly used Mel frequency cepstralcoefficients (MFCC), and builds a real-time distributed speech recognition front-endon the DSP platform.MFCC performs well and has low computation complexity, however, the DCT init has nothing to do with actual data and is quite redundant, resulting in poorrobustness of speech recognition systems in noisy environments. In order to overcomethis shortcoming, the thesis uses data-driven ICA method to replace the DCTprocedure. In ICA, the method based on symmetric orthogonalization is adopted toreduce the accumulated errors and can compute parallelly resulting in higherestimating speed. At the same time, the optimal non-linear functions and theircoefficients are fixed in the thesis. Experiments under a variety of noises andSNRs(Signal Noise Ratio) show that the average recognition rate of the improvedmethod is6.17%higher than that of MFCC, especially under low SNRs, so the newmethod improves the robustness of the speech recognition systems greatly.Based on the proposed improved MFCC method, the thesis builds a distributedspeech recognition front-end on the DSP platform for real-time extracting theimproved features. Because ICA is a kind of statistical method, it needs a lot of actualdata to train the statistical model, that is, the demixing matrix. Using an offlinemethod of calculation, the thesis computes the demixing matrix on the PC, anddirectly applies it to feature extraction in the distributed speech recognition front-end,improving the robustness of the features without increasing the complexity. This is ofgreat significance to practical application of distributed speech recognition.