What is not in the public record¶

Everything on the previous pages came from documents anyone can read: the Multi-Band Excitation papers, a set of expired US patents, and the JARL D-STAR system specification. Read together, they describe the MBE model completely enough that you could build an MBE vocoder from them.

They do not describe the vocoder in your radio.

This page is about that difference: a map of the boundary between what is documented and what is not, and of what independent measurement of real hardware can and cannot tell you about the far side of it.

What the public record actually gives you¶

More than most people assume. The expired patents are unusually specific, because a patent must enable the invention.

• The model itself. Speech as a spectral envelope multiplied by an excitation spectrum, with an independent voiced/unvoiced decision per frequency band rather than one per frame. Griffin & Lim 1988
• Pitch and voicing estimation, including the nonlinear band-preprocessing trick that regenerates energy at a missing fundamental, and the hybrid frequency-domain/time-domain estimator with pitch-multiple search and voicing smoothing. US 5,715,365; US 5,826,222; US 5,216,747
• Spectral amplitude estimation, computed identically for voiced and unvoiced harmonics — and, in one case, an actual tabulated analysis window printed coefficient by coefficient. US 5,754,974, Table 1
• Predictive quantization: predict this frame's log-amplitudes from the last frame's reconstructed ones, interpolate across a changing harmonic count, quantize the residual through a block DCT, subtract the mean so a gain error cannot propagate, and apply a decay factor below unity so channel errors die out. US 5,630,011
• The channel side: bits ranked by error sensitivity, Golay and Hamming codes assigned accordingly, interleaving against burst errors, a pseudo-random scrambling sequence keyed from the strongest-protected bits so the decoder can detect uncorrectable errors, and frame-repeat or mute on failure. This patent even prints a non-uniform gain quantizer codebook and a table of DCT step sizes. US 5,870,405
• Regenerated-phase synthesis, the thing that separates a natural MBE decoder from a buzzy one: voiced phases derived from the log spectral magnitudes through an antisymmetric kernel, with the kernel and its scaling constant given explicitly. US 5,701,390
• The D-STAR channel layer: 4800 bps GMSK, 96 bits per 20 ms frame, split as 72 bits of voice frame plus 24 bits of data, with the voice budget described as 2400 bps of voice and 1200 bps of FEC. JARL D-STAR system specification

That is a lot. It is genuinely enough to build a working MBE codec.

The gaps¶

The problem is that the public record describes a family of systems, and a shipping product is one specific member of it, chosen and tuned. The documents stop exactly where the choices begin.

The bit budget does not line up. The expired patents do give bit allocations — one of them allocates a 3.6 kbps system as a fundamental frequency, a set of band voicing bits, and the remainder to magnitudes US 5,701,390. But in that system the entire 3.6 kbps is spent on parameters, with forward error correction layered on top at a higher rate — 3.6 kbps of parameters carried in a 4.8 kbps channel. D-STAR inverts that: its 3600 bps voice frame already includes 1200 bps of FEC JARL, leaving 2400 bps of parameters — two thirds of the bits, not all of them. So whatever allocation the deployed D-STAR codec uses, it is not any of the allocations printed in the expired patents. No public document states what it is.

The band splitting is unfixed. The public sources variously describe sixteen analysis channels mapped to eleven voicing bands US 5,715,365, eight voicing bands US 5,701,390, and a range of three to fifteen bands quoted as context US 5,649,050. All are consistent with the model. Only one of them is in your radio, and the papers do not say which.

Free parameters are given as shapes, not values. The predictor decay is specified as "a factor below one, chosen as a function of the harmonic count" US 5,630,011, and elsewhere as adaptive US 5,870,405. That is a design constraint, not a number. The same is true of voicing thresholds, pitch-tracking penalties, and the error-rate thresholds that trigger frame repeat or mute.

Nothing fixes absolute level. The patents describe the gain term structurally — a frame-mean of the log magnitudes, quantized non-uniformly, transmitted differentially. They do not tie that term to a physical reference. Two implementations can both be internally consistent, both faithful to the patents, and still disagree by a large constant on what a given input amplitude should encode to. Worse, the gain term admits more than one reasonable definition — the log of the mean energy and the mean of the log magnitudes are not the same quantity, and they diverge as the spectrum spreads out.

Codebooks, if any, are absent. If a deployed codec replaces the patents' scalar DCT-residual quantizers with trained vector-quantizer codebooks, those tables exist in no publication. A codebook is not an algorithm; there is nothing to enable and nothing to disclose.

Those gaps are visible in one picture. The frame-bit figure on the D-STAR frame on the air shows 72 real captured bits per frame and colours them by how often each position changes, because that is a measurement. It does not colour them by field, because the field map is the thing that is missing. The part of that figure worth looking at is the part that stays deliberately unlabelled.

What black-box measurement can establish¶

Quite a lot, if you are careful about what you claim.

Driving a purchased AMBE-3000 with known signals establishes end-to-end transfer properties: the level relationship between input and decoded output, the spectral tilt a signal acquires crossing the codec, the latency, and the behaviour at the extremes of its dynamic range.

It establishes dynamic behaviour that no document mentions. Our measurements found that the device's gain tracking ramps up over roughly twenty to forty milliseconds at word and phoneme onsets, a systematic level shortfall at attacks that the ear reads as natural articulation rather than as a fault. That is a repeatable property of the product, and it is nowhere in the literature.

It establishes the shape of convention mismatches. Cross-connect an independent implementation and the hardware in both directions: if each side's own loop is spectrally flat while the two cross paths show equal and opposite tilt, that is the signature of a single scalar convention each loop silently cancels. The measurement localises the disagreement to one recursive parameter and constrains its value; it does not tell you what that parameter is called inside the device, or whether the device even has one.

And it establishes negative results, which is where most of the value is. A hypothesis that fits one probe and fails the cross-path test is refuted, cheaply. Several of ours were.

What it cannot establish¶

Measurement observes behaviour and therefore can never distinguish between implementations that behave identically. It cannot recover a codebook: probing a table tells you what a given index reconstructs to under one condition, not how the table was designed or what it contains elsewhere. It cannot see the encoder's internal decisions except as they survive into the emitted frame. It cannot separate "the AMBE full-rate definition" from "this chip, this firmware revision" — a second device from a different production era might not agree, and we have not tested that. And a model that fits every measurement you have taken is still only the simplest explanation for the measurements you took.

What remains unknown¶

Stated plainly, as open questions rather than as secrets:

1. The exact gain convention. We can measure the offset between an independent implementation and the hardware, and we can compensate for it empirically. We cannot yet derive it from first principles, and the shape of the residual error suggests the definition itself differs, not just a constant.
2. The parameter bit allocation of the deployed D-STAR full-rate frame. Not published, and not the same as any patent allocation.
3. The band-splitting actually used, and the thresholds governing the voicing decision.
4. Whether the decoder's spectral post-processing is on by default, and with what strength.
5. Whether "AMBE full-rate" is one stable definition across DVSI product generations, or a family that drifted.

What this page will never contain¶

There is a version of this chapter that lists numbers. This is not it.

This project only publishes claims traceable to an expired patent, a published paper or public specification, or its own measurement of purchased hardware. A number whose only provenance is somebody else's reverse-engineered implementation fails that test, so it is not here — several candidate claims were dropped from this page on exactly that ground.

That constraint turns out to cost little, because the shape of a gap is the interesting part. Knowing that the gain convention is ambiguous, and why two defensible definitions diverge, is worth more to someone building an interoperable codec than a constant that happens to work on one device. The constants are findable by anyone with a hundred-dollar stick and patience. The map is what takes time.

Next: Listen: real hardware examples. Eight sentences through a real AMBE-3000, with something specific to listen for in each. Previously: Synthesis: rebuilding the voice.

Last reviewed: 2026-08-10.