Research Of Speech Enhancement Algorithm Based On Dictionary Learning

The speech is inevitably polluted by the noise during communication, especially under strong background noise, which not only affects the semantic identification and reduce the communication efficiency.what’s more, a long time noisy speech can also affect one’s hearing and emotion, so an receiver equiped with noise supress session is quite essential. Speech enhancement are widely used in human-computer interaction systems, such as interactive, real-time translation, smart homes, and so on.Speech enhancement is an important branch of digital signal processing. Its main function is to suppress the noise and improve the intelligibility and intelligibility of speech. A big challenge in speech enhancement algorithm is the extracted speech under strong noise still remains some residual in the non-speech segments, makes it difficult for speech recognition and information delivery. Aimming at solving the above problems, a robust algorithm based on dictionaries is presented. The research is divided into three parts.1. An improved spectrum subtraction is proposed. Instead of short time energy and short-time zero-crossing rate, short-time autocorrelation and frequency variance is used in threshold settings of voice activity detection, which improves the accuracy of the noise spectral estimation. Experimental results show that the new method performs better on the SNR and PESQ assessment.2. Pure Speech could be very sparse in an appropriate transform-domain. The dictionary learning algorithm is introduced in speech enhancement, it’s consist of two parts, the tracking algorithm and the dictionary training algorithm. The later is optimized by applying the LS-OMP algorithm, which proves to be efficient.3. A comparative summary of the improved dictionary learning algorithm and the classical enhancement algorithm is made, the experimental results show that the proposed algorithm performs betteer than classical unsupervised algorithms.