| The speech recognition system has achieved high recognition accuracy under the ideal laboratory conditions. But in real application, the noise reduces the recognition performance. To improve the performance of the speech recognition system, especially under low SNR environment, this thesis is focused on speech endpoint detection and rejection to optimize the two important parts of a speech recognition system. The main work of this thesis is introduced as follows:(1) The structure of speech recognition system is analyzed, including preprocessing, feature extraction and acoustic modeling. The main work of preprocessing is pre-emphasis, framing/windowing and endpoint detection. Mel frequency cepstral coefficient (MFCC) and corresponding first-order differential cepstral coefficients is selected as the feature parameters and modeled by Hidden Markov model.(2) The endpoint detection algorithm based on band-partitioning spectral energy-entropy is realized and the endpoint detection module is optimized. Band-partitioning spectral energy-entropy parameter can present speech signal feature and is robust to different kind of noise. The test results show that endpoint detection based on band-partitioning spectral energy-entropy is better than that based on dual-threshold. The performance of the speech recognition system is improved significantly, especially under low SNR environment.(3) The rejection module is added to the speech recognition system. The rejection studies three aspects:short speech length, low SNR and possible recognition error. And the rejection algorithm based on likelihood difference is further studied. The recognition accuracy is improved significantly through the rejection of short speech length, low SNR or possible recognition error, and the instantaneity of the system is guaranteed with no extra systematic complexity.(4) Through a large number of experiments on MATALB platform, the validity of the endpoint detection module and rejection module is verified. Transplant the optimization part to C platform and realize the optimization of the real-time speech recognition system. The test results show that the recognition performance is improved significantly after optimization. |
