Voice Activity Detection Using an Improved Unvoiced Feature Normalization Process in Noisy Environments

Noise-elimination technology is used to eliminate noise, including environmental noise, from voice signals in order to increase voice recognition rates. Noise estimation is the most important factor in noise-elimination technology. One of the effective estimation methods is voice activity detection, which is based on the statistical properties of noise and voice. This method is a way of estimating noise using the statistical properties of both noise and voice, which have an independent Gaussian distribution. In cases of severe differences in a statistical property, like white noise, the method is very reliable but limited to signals having a low signal-to-noise ratio (SNR) or having speech shape noise, which has statistical properties similar to voice signals. Methods to increase the voice recognition rate suffer from decreasing voice recognition performance due to distortion of the voice spectrum and to missing voice frames, because noise remains if there has been incorrect estimation of the noise. Degradation in voice recognition performance emerges in the differences between the model training environment and the voice recognition environment. In order to decrease environmental discordance, various silence feature normalization methods are used. Existing silence feature normalization suffers from degradation of recognition performance because the classification accuracy for the voiced and unvoiced signals decreases by an increasing energy level in the silence section of a low SNR. This paper proposes a robust voice characteristic detection method for noisy environments using feature extraction and unvoiced feature normalization for a classification relative to the voiced and unvoiced signals. The suggested method constitutes a model for recognition by extracting the characteristics for classification of the voiced and unvoiced signals in a high SNR environment. Also, the model affects noise for voice characteristics less, and recognition performance improves by using the Cepstrum feature distribution property of voiced and unvoiced signals with a low SNR. The model was checked for its ability to improve recognition performance relative to the existing method based on recognition experiment results.