SDR in Pure Go, Part 12: Voice Coding — IMBE, AMBE+2 & MBE

By Matt Cheramie June 14, 2026

Part 12 of SDR Internals. Digital voice isn’t recorded audio — it’s a model of your voice, a few kilobits per second. This post covers the vocoders that reconstruct speech and the registry that makes them pluggable.

In this post

What a vocoder is and why digital radio can’t work without one.
GopherTrunk’s pure-Go IMBE (P25 Phase 1) and AMBE+2 (P25 Phase 2, DMR) decoders.
The shared MBE synthesis core and the Vocoder registry — a plugin + DRY design.

What a vocoder does

A vocoder compresses speech to a few kilobits per second by modeling how speech is produced — pitch, voicing, and a spectral envelope — rather than recording the waveform. The radio sends those model parameters; the receiver resynthesizes speech from them. Every digital voice mode depends on one. (digital voice)

GopherTrunk implements two in pure Go:

IMBE — 4.4 kbps, P25 Phase 1. (reference)
AMBE+2 — 2.4 kbps, P25 Phase 2, DMR, NXDN. (reference)

Both belong to the Multi-Band Excitation family — and that shared lineage is the key to the design.

How GopherTrunk implements it in Go

IMBE and AMBE+2 differ in how they unpack and error-correct their on-air bits, but they synthesize speech almost identically. So GopherTrunk factors the synthesis into a shared internal/voice/mbe core, and the two codecs become thin front-ends:

internal/voice/imbe — unpacks 144-bit IMBE frames, runs Golay/Hamming FEC, de-interleaves, and produces MBE parameters.
internal/voice/ambe2 — unpacks AMBE+2 frames, runs BPTC FEC and codebook lookups, and produces MBE parameters.
internal/voice/mbe — takes those parameters and synthesizes 8 kHz PCM: voiced-harmonic generation, unvoiced FFT excitation with overlap-add, spectral enhancement, and per-frame AGC.

Every vocoder satisfies one interface:

// internal/voice/vocoder.go (shape)
type Vocoder interface {
    Name() string
    FrameSize() int
    Decode(frame []byte) ([]int16, error)
    Reset()
    Close() error
}

…and registers itself with a factory registry, exactly like the SDR drivers from Part 2:

var DefaultRegistry = NewRegistry() // name -> VocoderFactory
// imbe/register.go and ambe2/register.go call Register() at init()

The design principle: plugin registry + shared-core DRY

Two principles combine. The plugin registry lets the engine pick a vocoder by name without importing it. And DRY — don’t repeat yourself — drives the shared mbe core, so the genuinely-common synthesis math is written once.

How that principle shaped the Go code

One synthesis engine, many front-ends. Because IMBE and AMBE+2 both reduce to MBE parameters, the hard DSP — harmonic synthesis, overlap-add, enhancement — lives in a single tested package. A future Codec2 or hardware vocoder reuses it.
Per-call instances, no shared state. Each active call gets its own Vocoder instance with its own Reset()-able state (cross-frame prediction needs memory), so two simultaneous calls never interfere.
Backends swap by name. The protocol decoder asks the registry for “imbe” or “ambe2”; the planned DVSI hardware backend slots in behind the same interface under a build tag, with zero changes to the decoders.
A NullVocoder always exists. Registering a silence vocoder means the pipeline always has a valid backend, so an unsupported codec degrades to silence instead of crashing.

Where this goes next

Vocoders are a deep, fascinating topic — the source-filter model, how MBE splits the spectrum into voiced and unvoiced bands, and the patent history that makes a clean-room Go implementation valuable. A future “Digital Voice in Go” series can dissect a single frame bit by bit. Next, we wire these audio samples into recordings and live streams.

FAQ

Why are IMBE and AMBE+2 implemented separately if they share a core? Their bit-packing and FEC differ completely — only the synthesis is shared. The split front-ends handle the on-air differences; the common mbe package handles what’s genuinely identical.

Is implementing AMBE in Go a patent issue? Re-implementing a codec in a new language doesn’t change its patent status. GopherTrunk provides the decoders; whether a given codec is encumbered depends on your jurisdiction. IMBE’s patents have expired; AMBE+2’s situation varies.

Why 8 kHz mono PCM output? That’s the native rate vocoders model speech at — 160 samples per 20 ms frame. It matches telephone-quality voice and keeps recordings small.

Part 12 of 14 · ← Part 11 · Next → Part 13: Recording, composition & streaming