Mandarin Digital Speech Recognition System Development Based On The Noise Environment

Noise robust for speech recognition system is key of speech recognition utility, and is development hotspot and difficulty of the speech recognition today. It is developed mostly that the speech detect algorithm, the speech enhancement technology and the feature extraction method in speech signal process in the paper in order to robust of the noise in the actual environment, it is mostly development as follows:1 .Traditional method of solving speech detect is short-time energy and zero-crossings rate, it is imposed that linking speech entropy with short-time energy, and improvement decision in order to speech detect and segmentation. The experiment results that the method can segment speech boundary exactly when the traditional method is no good use for this.2. It is discussed that wiener, spectral subtraction based on the noise energy spectral estimate and wavelet denoise method. Aim at elimination music background noise, it is imposed that adding weight function which make use of short-time energy and zeros-crossing rate denoise speech secondly that is processed by the spectrum subtraction based on the noise energy spectrum estimate.3. It is deeply analyzed and studied that the noise-restrained parameters, and is introduced that short-time energy, one-order, second-order difference and Cepstrum Mean Subtraction in the Mel Frequency Cepstrum Coefficient. It is imposed that make use of combined feature coefficient to enhance parameters antinoise.4. It is developed HMM model and discussed the state number of Markov chains, choosed of original value and the scale of trainging-set, the Gauss mixture number, etc. In addition, it is discussed to introduce noise transcendent knowledge, namely extend noisy to speech example database in order to advance in model robust for noise.Adopting hereinafter result of experiment analyse, it is realized that non-special , non-fixed-length madarin digital string speech recognition system based on the Continuous Density Hidden Markov Method in the actual system emulation, and estimated system performance from antinoise, recognition rate and error recognition rate for length etc.