How AMBE Works¶

A D-STAR radio takes a fifth of a second of your voice, throws the waveform away, and transmits nine bytes. This site is about what is in those nine bytes, how they get chosen, and how a decoder that has never heard you turns them back into a person talking.

Every claim here names its source in the sentence that makes it. The sources are US patents, the published Multi-Band Excitation literature, the JARL D-STAR system specification, public DVSI product documentation, and measurements this project made against a vocoder chip bought at retail. Nothing comes from anybody's codec source code.

Start with the picture¶

The whole codec follows from one observation: a voiced sound is a stack of harmonics of a single fundamental, and how many of those harmonics fit inside the coded band depends on how deep the voice is. Drag the slider. Watch the count in the readout fall from sixty-odd to nine.

That count is the number of amplitudes the coder has to describe, and the bit budget does not grow to meet it. Almost every design decision in AMBE is a consequence of that one squeeze. What AMBE actually is picks the argument up from there.

What this site is¶

An explanation, written for someone who wants to know what is actually happening inside a digital voice link rather than to be reassured that it works.

It assumes you are comfortable with a spectrum plot and the idea of a Fourier transform. It does not assume you have read a speech coding textbook. Where a patent states something in its own words, the patent is quoted and cited, so you can go and disagree with the reading.

Three kinds of thing appear on these pages, and they are labelled so you can tell them apart:

• Public teaching. The model, the analysis, the quantization structure and the synthesis rules, taken from the patents and from the 1988 papers. Cited by number.
• Specification. The over-the-air format, taken from the JARL D-STAR system specification. Cited by section.
• Measurement. Waveforms, spectra, pitch tracks, band voicing and captured channel frames from a DVSI AMBE-3000 operated as a black box, including the data behind the figures.

What this site is not¶

There is no working AMBE encoder or decoder in this repository, in any language. That is a constraint the project holds itself to, not an accident of scope.

Code in these pages is pseudocode, printed as plain text, with the constants, window definitions and band-edge arithmetic deliberately absent. It will not run. There are no quantizer codebooks and no bit-packing routines. The interactive figures draw precomputed JSON captured from hardware; they perform no pitch estimation, no voicing decision, no quantization and no synthesis.

The subject here is full-rate AMBE and IMBE, the codec D-STAR uses. AMBE+2 and the half-rate codec are covered at the level of history and patent status rather than mechanism, because they are a different codec. See the patent landscape.

This project is independent of Digital Voice Systems, Inc. AMBE, AMBE+ and AMBE+2 are trademarks of DVSI, used here only to identify the technology under discussion. Nothing on this site is affiliated with, sponsored by, endorsed by or approved by them, and nothing on the patent page is legal advice.

The reading path¶

The nine chapters are written to be read in order. Each one ends by handing the next one a problem it has to solve.

1. What AMBE actually is sets up the bit budget, shows why a single voiced/unvoiced decision per frame produces buzz, and introduces the multi-band model that fixes it. Start here.
2. Analysis: pitch, voicing, amplitudes is how a 20 ms window becomes one pitch, eight flags and a list of amplitudes. Analysis by synthesis, and why pitch is a search rather than a measurement.
3. Turning measurements into bits spends the budget: log-domain amplitudes, prediction from the previous frame, codebooks, and a bit allocation that has to change with the pitch.
4. The D-STAR frame on the air wraps the result in error correction and clocks it onto a carrier. Unequal protection, a descrambling key chained to the most important field, and the data channel riding alongside your voice.
5. Synthesis: rebuilding the voice is the far end: an oscillator bank, shaped noise, and the phase that was never transmitted and has to be invented fifty times a second.
6. What is not in the public record draws the boundary. What the documents give you, what a black box on a bench can settle, and the five questions that are still open.
7. Listen: real hardware examples is eight sentences in two voices, before and after a real AMBE-3000, with something specific to listen for in each pair.
8. The patent landscape is the dates: which patents taught what, which expired when, and which two are still live.
9. Sources and method lists every source, states what was deliberately excluded and why, and tells you how to check any of it.

In a hurry? Read chapter 1, then listen to a clip pair, then come back for the rest. If you arrived for the legal question, chapters 8 and 9 stand on their own.

Corrections¶

A page that is wrong and cited is more useful than a page that is vague and safe, because the citation is what lets someone catch the error. If a claim here is not supported by the source attached to it, please open an issue naming the page and the primary document that contradicts it.

Prose and figures are licensed CC BY 4.0; the tooling is MIT. Reuse either, with attribution.

Last reviewed: 2026-08-10.