DMR End to End, Part 11: AMBE+2 Voice & the Silence-Frame Bug

By Matt Cheramie · September 27, 2026

Part 11 of DMR End to End, a 14-part deep dive that follows the world’s most widely deployed digital PMR protocol through GopherTrunk — from a 4FSK carrier to two simultaneous calls, direct-mode handhelds, and decrypted Enhanced Privacy voice. Part 10 left a wideband tap producing clean superframes. This part follows them into the vocoder — the AMBE+2 3600×2450 decoder in internal/voice/ambe2 — and to the “computer voice” report of #644. Two of everything, once more: two mbelib lineages to diff against, two kinds of frame the decoder must tell apart, and two calibrations that were never in the vocoder core.

TL;DR: An earlier repo note blamed DMR’s rough audio on a “high-band deficit in unpackParams2450”. A frame-by-frame diff of the committed #644 clip (testdata/dmr-voice.raw, 378 frames) against szechyjs/mbelib 1.3.0 and mbelib-neo refuted it: Tl / Vl / L / w0 match to 1e-5 on all 201 voice frames. The defect was the other 177 — AMBE+2 silence frames (b0 124/125) that both references decode as an all-unvoiced fixed-model frame (w0 = 2π/32, L = 14, the frame’s own ΔΓ / PRBA / HOC) while GopherTrunk emitted digital silence and reset the gain predictor, so every onset after a pause decoded at gamma = ΔΓ + 0 instead of ΔΓ + 0.5·γ_prev — frame 95: 4.37 vs 8.40, ≈ 24 dB low. Fixed to the reference behaviour, pinned by params2450_silence_test.go. Two calibrations followed, measured against dsd-neo: ShapeUnvoicedSpectrumGain with DefaultUnvoicedGain = 5.49 (unvoiced bands were 13–15 dB low) and a first-order tilt_hz = 450 high-pass. LSD to dsd-neo: DMR male 4.9 → 2.1 dB. Still open: AMBE 2450 male voiced harmonics −3..−6 dB.

Key takeaways

• A refuted note is worth recording. “The deficit is in the 2450 tables” was plausible from whole-file band fractions and wrong; a per-frame diff against an independent decoder settled it.
• Silence frames are not silence. b0 124/125 carries a gain delta and an envelope; the references synthesise it and carry the predictor through. Zeroing prevGamma instead made every onset 24 dB quiet.
• Calibrate the band, not the ear. Per-harmonic DFT, band fractions and log-spectral distance against dsd-neo are the metrics; the shipped WAVs hid all of it under enhance, normalize and warm.
• Compare voice frames only. The references disagree on the silence model itself (L = 14 vs 15), so whole-file spectra depend on which you pick.

Cheat sheet

Concern	What it does	Where it lives
2450 unpack	b0..b8 scattered layout → dmr* tables → PRBA/HOC inverse DCTs → Tl	internal/voice/ambe2/params2450.go (unpackParams2450), tables2450.go
Silence frame	b0 124/125 ⇒ w0 = 2π/32, L = 14, all-unvoiced, own ΔΓ / Tl; SilenceFrame flag	params2450.go, params.go (Params.SilenceFrame)
Gain predictor	gamma = ΔΓ + 0.5·prevGamma, carried through pauses	decoder.go (prevGamma, foldGammaIntoTl)
Reference pins	literal frames vs mbelib cur_mp values	params2450_silence_test.go (TestDecodeDMRSampleGammaTracksReference)
Unvoiced band level	per-band power = gain·Ml²/2; 5.49 = mbelib’s 3-cosine level	internal/voice/mbe/synth_unvoiced.go (ShapeUnvoicedSpectrumGain, DefaultUnvoicedGain)
Radio tilt	first-order HPF, bilinear pre-warped, default 450 Hz	mbe/enhancer.go (radioTilt, EnhancerConfig.TiltHz)
Operator knobs	recordings.unvoiced_gain, recordings.enhance.tilt_hz, recordings.mbe_files	config.example.yaml, docs/vocoders.md

In this post

• From superframe to 49 bits — what reaches the vocoder, and which frames are not voice.
• The note that was refuted — the per-frame diff against two mbelib lineages.
• Silence frames are not silence — the predictor reset and the 24 dB onset.
• Calibrating the unvoiced band and the tilt — dsd-neo as the measured reference.
• Checking your own frames — .amb sidecars, DSD-FME’s 8 kHz header, decode.

From superframe to 49 bits

Part 3 sliced the voice superframe; each burst carries three 72-bit AMBE+2 frames, and the Golay(23,12) + descramble stage in internal/radio/dmr/voice reduces each to 49 bits (Voice Coding Part 8 covers that boundary). The DMR variant is ambe2-dmr: the 3600×2450 rate whose bit positions and codebooks (generated from mbelib’s ambe3600x2450_const.h) differ from the 3600×2400 default of Voice Coding Part 7. Everything after unpackParams2450 — the gamma fold, mbe.PredictLog2Ml, §6.2 enhancement, synthesis — is shared with the base decoder and IMBE.

In the 2450 frame, b0 is not always a pitch: 0..119 index dmrW0table, 120..123 mark an erasure, 126..127 a tone frame, and 124..125 an AMBE+2 silence frame — the subject of this part.

The note that was refuted

The #644 report was that DMR voice “sounds awful — unnatural, computer voice”, and an early measurement seemed to pin it. Against the operator’s DSD-FME decode of the same .amb frames, GopherTrunk’s ambe2-dmr output had 4–10× less energy above 1 kHz. The base ambe2 decoder shares the synthesis and gave mbelib-like highs, so the note concluded the deficit was “purely in unpackParams2450” and would need a frame-by-frame reference before anyone touched the tables.

That diff was done, and it refuted the note. szechyjs/mbelib 1.3.0 and arancormonk/mbelib-neo were built from source and made to dump cur_mp per frame over the committed 378-frame clip; on every one of the 201 voice frames the 2450 unpack’s Tl, Vl, L and w0 match mbelib to 1e-5 and the amplitude prediction within ~1 dB per band. TestUnpack2450TlMatchesMbelibFrame0 pins a literal on-air frame (b0b9243e0bc180, L = 37) against mbe_decodeAmbe2450Parms’s Tl — the independent-reference vector the self-consistent trap demands, since a round-trip could never catch table drift. Do not chase the 2450 tables. On voice frames only, GopherTrunk is not high-band deficient against mbelib-neo — the whole-file fractions were confounded by the other 177 frames.

Silence frames are not silence

Those 177 frames are AMBE+2 silence frames — 47 % of the clip, in runs of 44 and 63 (≈ 1.3 s) between utterances. Both references decode them as an all-unvoiced frame with a fixed model — w0 = 2π/32, L = 14 — carrying the frame’s own gain delta and PRBA/HOC envelope, then synthesise them like any other: the background plays at its level and the cross-frame state carries through. GopherTrunk short-circuited them to digital silence and reset the predictor:

// internal/voice/ambe2/params2450.go (shape)
silenceFrame := b0 == 124 || b0 == 125
if b0 >= 120 && !silenceFrame { /* erasure / tone → Silent */ }
if silenceFrame {
    f0, L = 1.0/32, 14 // mbelib fixed model: w0 = 2π/32, L = 14, all unvoiced
} else {
    f0, L = dmrW0table[b0], int(dmrLtable[b0])
}
// …b2 → p.DeltaGamma; PRBA/HOC → Tl, exactly as for voice

The gain is the cross-frame quantity AMBE+2 has and IMBE lacks: gamma = p.DeltaGamma + 0.5*d.prevGamma. Zeroing prevGamma on every silence frame meant the first voice frame after every pause decoded with gamma = ΔΓ + 0 — frame 95 at 4.37 against the references’ 8.40, frame 308 at 5.78 against 10.03: ≈ 24 dB low, on every onset, with the pauses hard-gated between. That is the “computer voice”, and none of it was in the tables.

The references carry the predictor through a silence run; the old decoder zeroed it, so each utterance began four log2 units below the reference and rose from there.

The fix is the reference behaviour: Params.SilenceFrame marks the frame, the unpack builds the fixed model with the frame’s own DeltaGamma and Tl, and the decoder synthesises it through the normal path, predictor intact. TestUnpack2450SilenceFrameIsUnvoicedNoiseFrame pins the literal frame 19 (f84902a06c3d80) and fails against the old Params{Silent: true}; TestDecodeDMRSampleGammaTracksReference checks prevGamma at nine named frames against mbelib’s cur_mp->gamma to 2e-3, and mbelib-neo agrees.

One thing the diff did not settle: the references disagree on the silence model itself — mbelib-neo/JMBE uses L = 15 and a different w0; mbelib and DSD-FME’s lwvmobile fork use 2π/32 and L = 14 — and GopherTrunk follows the DSD-FME lineage the reporter compares against. Whole-file band fractions depend on that, so spectral comparisons must be made on voice frames only.

Calibrating the unvoiced band and the tilt

With the voice frames bit-identical, the remaining “sounds awful” was measured, not debugged: a per-harmonic DFT of GopherTrunk’s and dsd-neo’s decodes of the same 10 Sep .imb/.amb pairs. f₀ was identical and voiced IMBE harmonics matched within ±1 dB above 1.5 kHz. Two things did not, and neither was in the vocoder core.

Unvoiced bands were 6 dB (IMBE) to 13–15 dB (AMBE+2 2450) low. mbelib synthesises an unvoiced harmonic as three random-phase cosines ≈ 1.353·Ml each — about 2.745·Ml² per band, 5.49× the Ml²/2 a voiced harmonic carries. GopherTrunk’s §6.4 FFT-noise path scaled each bin by Ml. The fix normalises per band:

// internal/voice/mbe/synth_unvoiced.go (shape)
const LegacyUnvoicedGain  = -1   // old per-bin Ml scaling, kept for goldens
const DefaultUnvoicedGain = 5.49 // mbelib's 3-cosine level ≈ 2.745·Ml²
// gain ≥ 0: each unvoiced band's expected power is gain·Ml²/2, whatever
// its FFT-bin span; gain < 0 reproduces legacy byte-for-byte.
func ShapeUnvoicedSpectrumGain(spec []complex128, p Params, M *[57]float64, gain float64)

recordings.unvoiced_gain exposes it: 0 ⇒ the 5.49 default, 1 = the equal-power spec reading, < 0 = legacy. The raw decoders start legacy so their goldens hold; the recorder and gophertrunk decode set the default, and decode -legacy-synthesis opts out for A/B.

Below ~1.5 kHz GopherTrunk carried a smooth low-frequency excess — +4 dB at 500 Hz rising to +16 at 125 — the shape of a first-order high-pass, which is what dsd-neo runs by default (use_hpf_d=1) and what a handset does. radioTilt is that filter, bilinear-transformed with a pre-warped corner so the −3 dB point lands on fc at 8 kHz (the naive RC recurrence is ~1.2 dB off). Fitting voiced harmonics alone gives ~750 Hz; with the unvoiced level also calibrated, the corner minimising log-spectral distance to dsd-neo across all four pairs is ~450 Hz, so tilt_hz: 450 is the default.

The verdict is the number, not the ear: LSD to dsd-neo over 300–3400 Hz, P25 male 8.3 → 1.5 dB, DMR male 4.9 → 2.1 dB. What remains open is real: AMBE 2450 male voiced harmonics sit −3..−6 dB below dsd-neo above 1 kHz (IMBE doesn’t show it; the DMR female goes the other way), and that one is in the 2450 amplitude path — it needs the mbelib-neo per-frame diff. The operator’s warm_dmr_audio: true plus enhance LPF only made it worse — Voice Coding Part 10 has the PCM chain these knobs sit in.

Checking your own frames

The method that settled this is available to any operator, and Voice Coding Part 12 built the harness around it. recordings.mbe_files: true writes DSD-FME’s cookie-headed container next to each recording (.amb for DMR, .imb for P25 Phase 1), and DSD-FME decodes it directly. One quirk: DSD-FME’s -w writer stamps an 8 kHz header on 12 kHz synthesis, so the file plays ~1.5× slow. dsd-neo writes a correct 8 kHz WAV and is the reference the calibration was measured against:

for f in *.amb; do ./dsd-neo -o null -w "${f%.amb}_dsd.wav" -fs -r "$f"; done
gophertrunk decode -in call.raw -out call_gt.wav -vocoder ambe2-dmr   # recorder defaults
gophertrunk decode -in call.raw -out call_raw.wav -vocoder ambe2-dmr -legacy-synthesis

Then compare with band fractions, centroid or LSD — on voice frames — never by ear alone. The “gt-shipped” WAVs that started the report had enhance, loudness normalize and warm_dmr_audio stacked on top and hid every effect above.

How the diff shaped the Go code

• Silence is a flag, not a short-circuit. Params.SilenceFrame rides the normal voice path; only erasure and tone frames set Silent.
• Reference literals over round-trips. params2450_silence_test.go carries hex frames and cur_mp values from two independent decoders — the only shape that catches table or predictor drift.
• Calibration lives in the caller. LegacyUnvoicedGain keeps the raw decoders’ goldens byte-identical; the recorder and decode apply DefaultUnvoicedGain and the tilt per call.
• Every knob landed in config.example.yaml — unvoiced_gain and enhance.tilt_hz with their measured defaults.

Where this goes next

Clear voice now decodes as the references do. The next part is voice that was never meant to be clear: Part 12 follows DMR Enhanced Privacy from the PI header through RC4 keyed with key‖MI, the 256 dropped keystream bytes, and the embedded IV that names the next superframe — and why a Golay histogram was the first instrument, not a cipher theory.

FAQ

Why did GopherTrunk’s DMR audio sound like a “computer voice”? Not the vocoder tables: on the #644 clip every voice frame’s parameters match mbelib to 1e-5. The cause was AMBE+2 silence frames (b0 124/125), which the decoder rendered as digital silence while resetting the gain predictor, so each utterance after a pause started ≈ 24 dB low. They now decode as the references do.

What is an AMBE+2 silence frame? A 3600×2450 frame whose b0 is 124 or 125. mbelib and mbelib-neo both decode it as an all-unvoiced frame with a fixed model (w0 = 2π/32, L = 14 in the DSD-FME lineage) and the frame’s own ΔΓ / PRBA / HOC bits, then synthesise it — the background the radio encoded plays at its transmitted level.

Is the “2450 high-band deficit” real? No. That earlier note was refuted by a frame-by-frame diff against szechyjs/mbelib 1.3.0 and mbelib-neo: unpackParams2450 is bit-identical on every voice frame; the whole-file fractions that suggested it were confounded by 47 % silence frames. What remains open is a −3..−6 dB deficit in male voiced harmonics above 1 kHz on AMBE 2450 specifically.

What does recordings.unvoiced_gain do? It sets the unvoiced band power relative to a voiced harmonic of the same amplitude. 0 selects the default 5.49 — mbelib’s three-cosine level, 6–15 dB above GopherTrunk’s old scaling; 1 is the spec’s equal-power reading; a negative value restores legacy. The recorder and decode apply it per call.

How do I compare GopherTrunk’s DMR decode against DSD-FME? Set recordings.mbe_files: true for a .amb sidecar and decode it with DSD-FME or dsd-neo. DSD-FME’s -w stamps an 8 kHz header on 12 kHz audio, so relabel or resample first. Decode the .raw with gophertrunk decode -vocoder ambe2-dmr and compare band fractions or LSD on voice frames.

Series navigation

Part 11 of 14 · ← Part 10: Wideband DMR — Bin Edges & the Deaf Heal · Next → Part 12: Enhanced Privacy — RC4 & the IV That Names the Next Superframe