DSP Blockset

Analysis and Synthesis of Speech

In modern digital systems, a speech signal is represented in a digital format that is comprised of a sequence of binary bits. For storage and transmission applications, it is often desirable for the speech signal to be represented in as few bits as possible, while maintaining its perceptual quality.

In narrowband digital speech compression, digital speech signals are sampled at a rate of 8000 samples per second. Typically, each sample is represented by 8-bits. This corresponds to a bit rate of 64 kbits per second. Further compression is possible at the cost of quality. Most of the current low bit rate speech coders are based on the principle of linear predictive speech coding. The simplest implementation of this compression technique is presented in the linear prediction coefficient (LPC) Analysis and Synthesis of Speech demo. This topic describes this demo, which models the theory behind signal transmission:

1. Open the LPC Analysis and Synthesis of Speech demo by typing dsplpc at the MATLAB command line.
   
   

1. The input to this model is a human speech segment that is 0.5 second long. This model calculates the reflection coefficients of the speech segment and uses them to create the linear prediction analysis filter (lattice-structure). The model calculates the residual signal by filtering the preemphasized speech samples. The residual signal, which is the output of the analysis stage, is the low energy equivalent of the input signal and is easier to quantize. The blocks in the synthesis stage of the model use the known residual and reflection coefficients to synthesize the original signal.

2. Simulate this model.
3. Double-click the Original Signal and Processed Signal blocks and listen to both the original and the processed signal.

1. There is almost no difference between the two. This is because no quantization is used.

Suppose that you want to create a model that more accurately portrays what happens during cellular phone communication. A better approximation of a real-world system would involve the quantization of the residual and reflection coefficients before they are transmitted. For information on how to design a scalar quantizer to accomplish such a task, see Creating a Scalar Quantizer.

Buffering Delay and Initial Conditions

Creating a Scalar Quantizer