Crate mbelib_rs 

Pure Rust AMBE 3600×2400 voice codec decoder for D-STAR digital radio.

The AMBE (Advanced Multi-Band Excitation) 3600×2400 codec compresses speech at 3600 bits/second with 2400 bits of voice data and 1200 bits of forward error correction (FEC). It is the mandatory voice codec for the JARL D-STAR digital radio standard, used in all D-STAR transceivers and reflectors worldwide.

Each voice frame is 9 bytes (72 bits), transmitted at 50 frames per second (20 ms per frame). The codec models speech as a sum of harmonically related sinusoids, with each band independently classified as voiced or unvoiced.

Usage

use mbelib_rs::AmbeDecoder;

// Create one decoder per voice stream — it carries inter-frame state
// needed for delta decoding and phase-continuous synthesis.
let mut decoder = AmbeDecoder::new();

// Feed 9-byte AMBE frames from D-STAR VoiceFrame.ambe field.
let ambe_frame: [u8; 9] = [0; 9];
let pcm: [i16; 160] = decoder.decode_frame(&ambe_frame);

// Output: 160 samples at 8 kHz, 16-bit signed PCM (20 ms of audio).
assert_eq!(pcm.len(), 160);

Decode Pipeline

Each frame passes through these stages:

1. Bit unpacking — 72-bit frame → 4 FEC codeword bitplanes
2. Error correction — Golay(23,12) on C0 and C1 (3-error correction). AMBE 3600×2400 does not apply Hamming to C3; those 14 bits are copied verbatim into the parameter vector.
3. Demodulation — LFSR descrambling of C1 using C0 seed
4. Parameter extraction — 49 decoded bits → fundamental frequency, harmonic count, voiced/unvoiced decisions, spectral magnitudes. Frames with b0 in the erasure range (120..=123) or tone range (126..=127) trigger the same repeat/conceal path as Golay-C0 failures, since D-STAR does not use codec-level tone signaling.
5. Spectral enhancement — adaptive amplitude weighting for clarity
6. Adaptive smoothing — JMBE algorithms #111-116, gracefully damps spurious magnitudes/voicing decisions on noisy frames
7. Frame muting check — comfort noise on excessive errors or sustained repeat frames (JMBE-compatible)
8. Synthesis — voiced bands per-band cosine oscillators (with JMBE phase/amplitude interpolation for low harmonics) plus a single FFT-based unvoiced pass (JMBE algorithms #117-126)
9. Output conversion — float PCM → i16 with SIMD-vectorized gain and clamping

Modules

validation

Validation-only exposure of the internal quantize pipeline.

Structs

AmbeDecoder

Stateful AMBE 3600×2400 voice frame decoder.

AmbeEncoder

Top-level D-STAR AMBE 3600×2400 encoder.

EncoderBuffers

Per-stream encoder working buffers.

FftPlan

Per-stream FFT planning cache.

PitchEstimate

Result of a pitch-estimation pass on one 20 ms frame.

PitchTracker

Per-stream pitch-tracker state, matching OP25’s pitch_est member variables. All fields are prefixed prev by design — they’re the rolling history the look-back / cumulative-error tests consume.

SpectralAmplitudes

Per-frame spectral amplitudes.

VuvDecisions

Per-frame V/UV decision vector.

VuvState

Per-stream V/UV state carried across frames.

Enums

DecodeError

Errors that can occur during AMBE frame decoding.

Constants

MAX_BANDS

Maximum number of V/UV bands (12 per AMBE spec).

MAX_HARMONICS

Maximum number of harmonic amplitudes (matches IMBE/AMBE L).

Functions

analyze_frame

Ingest one 160-sample frame and run the front-end analysis.

compute_e_p

Compute IMBE’s E(p) detectability function for one 301-sample pitch-estimation buffer.

decode_trace

Inspection helper for golden-vector validation.

detect_vuv

Stateless one-shot V/UV convenience wrapper.

detect_vuv_and_sa

Integrated V/UV + spectral amplitude extraction.

extract_spectral_amplitudes

Extract harmonic amplitudes from the FFT half-spectrum.

pack_frame

Pack a 72-bit FEC-codeword array into a 9-byte AMBE wire frame.