Field Guide · algorithm

Also known as: MELP, MELPe, mixed-excitation linear prediction

MELP (Mixed-Excitation Linear Prediction) is a very-low-rate speech vocoder that improves on classic LPC by driving the vocal-tract filter with a mixed voiced/unvoiced excitation that varies across frequency, delivering intelligible speech at just 2400 or 1200 bps.1 Its enhanced revision, MELPe, is standardised as NATO STANAG 4591 and is the primary low-rate voice codec for military and secure HF/UHF communications.

pitch pulses noise per-bandmix band splitV/U weights LPC synth1/A(z) speech
MELP mixes pulse and noise excitation independently in each frequency band, so a frame can be voiced at low frequencies and noisy at high — cured of LPC's buzzy either/or.

How it works

Classic LPC-10 forces every frame to be either fully voiced (a pitch pulse train) or fully unvoiced (noise), which produces its notorious buzzy, mechanical timbre. MELP fixes this by enriching the excitation while keeping the same low-rate LPC spectral model:

  • Band-wise mixed excitation. The excitation spectrum is divided into frequency bands (five in the standard), and each is assigned a voicing strength blending a periodic pulse component with noise. Voiced speech with breathy or fricative high frequencies is represented naturally instead of being forced into one category.
  • Aperiodic pulses. A jitter flag lets the pulse train be slightly irregular, capturing the roughness of transitional and creaky voicing.
  • Adaptive spectral enhancement and a pulse-dispersion filter sharpen formants and smooth the synthetic buzz.
  • LPC envelope. As in all LPC coders, line-spectral-pair coefficients carry the vocal-tract shape; the encoder also transmits pitch, gain, and the band voicing flags — all packed into a 54-bit frame every 22.5 ms at 2400 bps.

The result is markedly more natural and robust in noise than LPC-10 at the same or lower rate — the reason it displaced older federal-standard vocoders.

Variants

The original MELP is US Federal Standard MIL-STD-3005 at 2400 bps. MELPe (enhanced MELP, STANAG 4591) adds a noise pre-processor, better analysis, and two extra rates — 1200 bps and 600 bps — while remaining interoperable with the 2400 bps mode. MELPe won the NATO selection over competitors and is embedded in tactical radios worldwide.

Relevance to SDR

MELP/MELPe is the vocoder of military and government HF/tactical voice, usually wrapped in strong encryption, rather than of the civilian trunked systems GopherTrunk targets — those use the multi-band-excitation family such as IMBE and AMBE+2. GopherTrunk does not decode MELPe (and where it appears the audio is typically encrypted anyway). It belongs on the map as the extreme low-rate, mixed-excitation cousin of the CELP and MBE families, showing how far speech can be compressed while staying intelligible.

Sources

  1. Mixed-excitation linear prediction — Wikipedia, for band-wise mixed excitation, bit rates, and STANAG 4591. ↩

See also