Linear Predictive Coding

Linear Predictive Coding (LPC ) is a process of the audio signal and speech processing, which is used, inter alia, for the audio data compression and voice recognition. Here, a simplified model of the human vocal tract is formed with which the language can be synthesized. Thus, a speech signal can only be described by the model parameters and excitation patterns. In comparison to the PCM representation this usually involves considerably less data. Usually, an additional error signal is transmitted to yet, which describes the difference between model prediction and the original signal.

The vocal tract model often consists of a vibration generator to mimic the vocal cords or the larynx, and a simple tube as a representation of the vocal tract downstream or articulation area (throat, mouth ). The vibration generator generates a periodic vibration, which can be modulated in their volume and pitch. The tube represents the resonant cavity of the throat and mouth and is used in cases of simple vowels usually as an adequate representation, while the representation of nasals branches of the tube ( nasal cavity ) would require that would be mathematically represent far more complex. Therefore, these sound shares are partly rejected and partly described by means of a residual signal.

A well-known example of the use of LPC for audio data compression is CELP, which compresses the residual signal using codebooks. The standardized method LPC 10e operates without codebooks and leads to a lower bit rate at the expense of sound quality.