DMR voice quality (AMBE+2)

Conventional DMR now decodes end to end, but the decoded voice — female voices especially — can sound rough. This page records what the decode chain does, a mbelib reconciliation review (what was ruled out and the one genuine divergence found), the operator levers you can pull today, and the calibration-gated path to a verified fix.

The short version: GopherTrunk’s default DMR decoder (internal/voice/ambe2) is a clean-room Go re-implementation of mbelib’s ambe3600x2450 path. mbelib is a reverse-engineered, approximate AMBE+2 decoder — it is known to sound worse than DVSI silicon, and worst on high-pitched (female) voices. So a large part of what you hear is inherent to the algorithm being mirrored, not a discrete bug. That said, one genuine divergence from the reference was found (below), and it is the prime suspect for the female-voice-specific roughness.

Why female voices are the hard case

A female voice has a higher fundamental ω₀, hence fewer harmonics L (internal/voice/ambe2/tables2450.go: L ranges 9…56, and the low-L rows are the high-pitch ones). Everything in the model that is indexed by harmonic position l relative to L therefore behaves differently for small L — most of the band sits in the “high band” (8l > L) where the spectral-amplitude enhancement acts, and the voiced/unvoiced and phase-dispersion decisions cover proportionally more of the voice. So a mis-tuned or mis-transcribed high-band step hits female voices hardest.

Reconciliation review (mbelib ambe3600x2450 + TIA-102.BABA §6.2)

Ruled OUT — these faithfully match the reference

  • Pitch / L / w₀ tables (tables2450.go) — generated from mbelib’s ambe3600x2450_const.h; the b0→(w₀, L) lookup matches.
  • Voiced/unvoiced decode (params2450.go:60-72). The jl = int(l·16·f0) index and the dmrVuv[b1][jl] lookup match mbelib. The clamp to [0,7] is a defined-behaviour guard; across the real f0 range jl spans the full 0…7 of the table (e.g. at the highest f0≈0.05, jl = 0.8·l reaches 7 only at l≈9), so there is no pathological V/UV collapse for female frames.
  • §6.3 voiced-phase de-buzz (decoder.go:synthFrame, ~627-664). mbelib does PHIl = PSIl for l ≤ L/4 and PHIl = PSIl + numUv·rand_phase()/L above it, with rand_phase() ∈ [-π, π]. GT draws (-1..1)·π · (numUv/L) for l > L/4 voiced harmonics — the same distribution and scaling. Faithful.

Found — the one genuine divergence (§6.2 spectral-amplitude enhancement)

internal/voice/mbe/enhance.go EnhanceAmplitudes is on the default decode path (ambe2/decoder.go:485, and imbe/decoder.go:552 for P25). Its per-harmonic weight ξ drops a factor that is present in both references the code cites (TIA-102.BABA §6.2 and mbelib’s mbe_spectralAmpEnhance):

mbelib / spec:   ξ = 0.96·π·num / ( ω0 · R_M0 · (R_M0² − R_M1²) )
GopherTrunk:     ξ = 0.96·num   / (       R_M0 · (R_M0² − R_M1²) )
                                   └─ missing the  π / ω0  factor ─┘
        num = R_M0² + R_M1² − 2·R_M0·R_M1·cos(ω0·l),   W_l = ξ^0.25, clamped [0.5, 1.2]

For DMR, ω₀ ∈ [~0.05, ~0.31] rad, so π/ω0 ∈ [~10, ~63]. GT’s ξ is therefore 10–63× smaller than the spec’s; after the ^0.25, GT’s high-band weights are ~1.8–2.8× lower, so they land on the 0.5 (attenuate) clamp where the spec would push toward the 1.2 (boost) clamp. Because the frame is then energy-renormalised (γ = √(R_M0/Σ)), the relative spectral tilt shifts energy out of the high band into the low band — a duller timbre, and, since small-L (female) frames have most of their harmonics in that high band, the effect is strongest on female voices. This matches the reported symptom.

enhance.go’s own header comment flags this as unfinished: “spec-tuning to bit-match mbelib output is part of step 5c (gain calibration).” So it is a known TODO, not a deliberate design choice.

The one-line correction (restore the spec/mbelib factor) is:

// internal/voice/mbe/enhance.go, inside EnhanceAmplitudes:
xi := 0.96 * math.Pi * num / (p.W0 * den)   // was: xi := 0.96 * num / den

Why this is not shipped as a default change yet. It is deliberately left for the calibration A/B, per this repo’s #764/#771 discipline:

  1. It cannot be verified to improve audio without a reference decode of the same frames — mbelib is itself approximate, and matching its (or the spec’s) number is not proof of better perceived quality.
  2. EnhanceAmplitudes is shared with the P25 IMBE path, so a blind change has blast radius beyond DMR.

The calibrate harness below is the gate: with a real reference WAV in place, the TestCompareAMBE2 cross-correlation will rise if this correction helps and fall if it hurts — decide from that, not from the formula.

A secondary, lower-confidence note: mbelib additionally multiplies the weight by √M_l (Wl = sqrtf(Ml[l]) · powf(...)), which the spec closed form does not. GT (correctly, per spec) omits it. Leave it omitted; it is a mbelib artefact, not part of the spec.

Levers you can pull today (no code change)

These are perceptual, not spec-exact, but cost nothing to try and can meaningfully warm up / brighten DMR audio while the calibration above is pending:

  • recordings.enhance — the opt-in output chain (EnhanceConfig). A high-shelf and a louder AGC target help intelligibility:

    recordings:
  write_raw: true          # also needed for calibration below
  enhance:
    enabled: true
    hpf_hz: 200            # trim rumble
    lpf_hz: 3400           # telephone band
    shelf_hz: 1600         # gentle presence
    shelf_db: 2            # dB of cut above shelf_hz (positive; <=0 disables)
    agc_target: 22000      # louder playback

  • ambe2-dmr-warm — the warm DMR vocoder variant (a high-shelf ToneTilt), selected by recordings.warm_dmr_audio: true (WarmDMRAudio). Softens the synthetic “edge” without touching the model decode.

Verified-fix path (calibration-first)

You already have the input side: every DMR call the daemon records with recordings.write_raw: true writes a <UTC>_<id>.raw sidecar of the packed AMBE+2 frames (7 bytes/frame) next to the .wav. Pick a female-voice call — the hard case — and:

  1. Reference-decode the same .raw through an independent AMBE+2 decoder:

    dsd-fme -r <call>.raw -o reference.wav      # 8 kHz mono 16-bit PCM

  2. Drop both files in and run the A/B (the test skips until the reference lands, so CI stays green today):

    cp <call>.raw    internal/voice/ambe2/testdata/dmr-voice.raw
cp reference.wav internal/voice/ambe2/testdata/dmr-voice-dsdfme.wav
go test ./internal/voice/calibrate/ -v -run TestCompareAMBE2

    Acceptance: |RMSRatioDb| < 3.0 (loudness) and PeakXcorr > 0.85 (waveform similarity). A clean RMS but low xcorr is the synthesis signature — i.e. the §6.2 enhancement above, not a gain knob.

  3. A/B the §6.2 correction. Note the baseline PeakXcorr, apply the one-line xi change, re-run. If xcorr rises (and DMR + P25 both hold or improve), ship it; if not, leave the default and keep the correction documented here.

See Voice calibration for the full harness reference (the cmd/voice-calibrate CLI, the IMBE path, and the AGC TargetPeak knob for the loudness half).

Do not commit real over-the-air public-safety recordings as fixtures. The calibrate testdata dirs are for local, operator-provided captures — the tests t.Skip when the files are absent, which is the intended state in the repo.