| The speaker recognition that distinguish the respect in the automatic identity has very important realistic meanings. At present, for instance, LPCC ,etc. adopted in characteristic analysis in the recognition system, all these methods are linear, only suitable for the steady linear time array, the signal of the speech is non-linear and non-steady, using these methods , should regard the signal of the speech as the steady signal in short-term ,if so, it will lose a lot of containing essential important information. Because the speech signal is non-linear and non-steady, on this basis ,this text has proposed that one kind processing method HHT that can reflect effectively of the non-linear characteristic of signal (Hilbert-Huang transform), and put forward a kind of HF characteristic that can reflect speaker's personal characteristic, proved as a kind of new characteristic of speaker's recognition system.This text set up a speaker verification system on the basis of HF characteristic that draw from HHT, all speech signal adopted come from YOHO speaker verification speech corpus. The YOHO Speaker Verification Corpus supports development, training and testing of speaker verification systems that use limited vocabulary, free-text input. Enroll stage: Adopting 138 speaker,for each speaker,there are 4 sessions and each one sessions has 10 utterances; Verify stage: Adopting 138 speaker,for each speaker ,there are 10 sessions and each one sessions has 4 utterances. Testing the speaker verification system , have made better result.The comparative analysis of the experimental data is explained: the accuracy of HF parameter drawing base on Non-linear theoretical foundation greater than LPCC. This kind of method is compared with LPCC, reaching same recognitionment result , drawing HF characteristic, the times of training is smaller than the times drawing LPCC parameter. So HF parameter based on HHT can be well applied to speaker recognition system, and have very broad application prospects. |
PDF Full Text Request