Speech compression techniques based on gaussian mixture model and wavelet analysis

الملخص EN

The basic goal of speech data compression is to reduce the bit rate for transmission and storage while either maintaining the original quality or providing an acceptable fidelity.

This article introduces two techniques for Arabic speech compression based on Gaussian Mixture Model (GMM) and Wavelet Transform (WT).

The performance evaluation of these techniques is carried out.

The Expectation Maximization (EM) is used to extract the GMM model.

A supervised learning technique is implementing based on EM algorithm to evaluate the model parameters of the probability density function (pdf).

The probability density function is used to build an optimized quantizer based on Lloyed algorithm.

Different compactly supported wavelet functions are investigated to get the proper wavelet function which should be used for Arabic speech compression.

The effect of increasing the level of decomposition and threshold values on the compression ratio are also investigated.

Different compression ratios at different signal to noise ratios are calculated.

The results show that the compression ratio becomes approximately constant after level 3.

The results also show that the Bior3.1 and Db10 wavelet functions give the highest compression ratio.

A comparison between the two techniques is given.

The speech quality after compression is measured using MOS ranking test.