edVirtus Reference

What Is Mixed Excitation Linear Prediction?

What Is MELP?

Mixed Excitation Linear Prediction (MELP) is a low-bit-rate speech coding technique that represents speech using a mathematical model of the human vocal tract rather than transmitting the speech waveform directly. Developed during the 1990s as an improvement over earlier Linear Predictive Coding (LPC) vocoders, MELP produces significantly more natural and intelligible speech while operating at bit rates as low as 2.4 kbps. It has been widely adopted in military, secure voice, and satellite communication systems where bandwidth is limited.

Like other vocoder technologies, MELP is based on the source-filter model of speech production. The speech signal is analysed to estimate the characteristics of the speaker's vocal tract together with the excitation that drives it. Instead of transmitting every speech sample, the encoder sends only a small set of parameters describing this model. At the receiving end, a speech synthesizer reconstructs an approximation of the original speech using the transmitted parameters.

The principal innovation of MELP is its use of mixed excitation. Earlier LPC vocoders typically assumed that speech was produced using either a periodic excitation for voiced sounds or random noise for unvoiced sounds. Real speech, however, often contains both periodic and noise-like components simultaneously. MELP models this behaviour more accurately by combining several excitation signals, producing much more natural-sounding speech, particularly for consonants and transitional speech sounds.

A useful analogy is recreating an orchestra using several different musical instruments rather than relying on a single instrument. Earlier vocoders often used only one type of excitation at a time, while MELP blends several excitation sources to produce a richer and more realistic result.

In addition to mixed excitation, MELP incorporates several other improvements over conventional LPC. These include adaptive spectral shaping, enhanced pitch modelling, improved gain estimation, and more accurate representation of aperiodic speech components. Together, these refinements significantly improve speech quality while maintaining very low transmission bit rates.

Because of its excellent performance at low data rates, MELP has been widely used in tactical military radio systems, secure telephony, satellite communications, and other bandwidth-constrained applications. It was adopted as a United States Federal Standard for secure voice communications and has been incorporated into numerous military communication systems worldwide.

It is important to distinguish MELP from Code-Excited Linear Prediction (CELP). Both are model-based speech coding techniques, but CELP employs a codebook of excitation signals to achieve higher speech quality at moderate bit rates, whereas MELP focuses on producing intelligible speech at extremely low bit rates using a mixed excitation model. Consequently, MELP is generally preferred where communication bandwidth is severely limited.

Today, MELP remains one of the most successful very-low-bit-rate speech coders. Although newer speech codecs provide higher quality at higher data rates, MELP continues to be valued for applications where bandwidth efficiency, robustness, and intelligibility are more important than natural speech quality. Its development marked a significant advance in speech coding and demonstrated how improved modelling of human speech production can greatly enhance communication performance at extremely low bit rates.