edVirtus Reference

What Is Multipulse Excitation?

What Is Multipulse Excitation (MPE)?

Multipulse Excitation (MPE) is a speech-coding technique that improves the quality of Linear Predictive Coding (LPC) by representing the excitation signal as a small number of carefully selected pulses within each speech frame. Rather than using a simple periodic pulse train or random noise, MPE identifies the amplitudes and positions of several excitation pulses that best reproduce the original speech. This results in significantly more natural-sounding speech while maintaining relatively low transmission bit rates.

Like other predictive speech coders, MPE is based on the source-filter model of speech production. The vocal tract is represented by a mathematical filter derived using Linear Predictive Coding, while the excitation signal represents the airflow produced by the vocal cords. The principal challenge is to represent this excitation efficiently without transmitting the complete speech waveform.

Earlier LPC systems generally used either a periodic pulse train for voiced sounds or random noise for unvoiced sounds. While this approach achieved low bit rates, the reconstructed speech often sounded synthetic and lacked naturalness. MPE overcomes this limitation by allowing several excitation pulses to occur within each analysis frame, with both their positions and amplitudes chosen to provide the closest possible approximation to the original speech.

A useful analogy is sketching a picture using only a small number of carefully placed dots. Rather than drawing every line, the artist positions a limited number of dots where they contribute most to the overall image. Similarly, MPE represents the excitation signal using only a limited number of strategically positioned pulses that capture the essential characteristics of the original speech.

During encoding, the speech signal is divided into short frames, typically 10–20 milliseconds long. The encoder analyses each frame and determines the optimum locations and amplitudes of the excitation pulses that minimise the difference between the reconstructed speech and the original speech. Only the pulse locations, pulse amplitudes, and LPC parameters need to be transmitted, resulting in efficient compression.

Compared with conventional LPC vocoders, MPE provides significantly better speech quality because it models the excitation signal more accurately. However, determining the optimum pulse positions requires considerable computational effort. This complexity ultimately encouraged the development of Code-Excited Linear Prediction (CELP), which achieves similar or better quality by selecting excitation sequences from a predefined codebook rather than searching for individual pulse positions directly.

One important application of MPE was the GSM Full-Rate speech codec, which employed a variant known as Regular Pulse Excitation with Long-Term Prediction (RPE-LTP). This codec became one of the first widely deployed digital mobile telephone speech coders, demonstrating the practical value of excitation-based predictive speech coding.

It is important to distinguish Multipulse Excitation (MPE) from Regular Pulse Excitation (RPE). In MPE, the excitation pulses may occur at arbitrary positions within the speech frame, allowing highly accurate modelling of the excitation signal. In RPE, the pulses are constrained to occur at regularly spaced positions, reducing computational complexity and transmission overhead at the expense of some modelling flexibility.

Today, MPE is regarded as an important milestone in the evolution of digital speech coding. Although it has largely been superseded by CELP, MELP, and newer speech codecs, its introduction demonstrated that improving the excitation model could greatly enhance speech quality without substantially increasing the transmission bit rate. Many of the concepts first explored in MPE continue to influence modern predictive speech-coding techniques.