Coverage & link robustness for voice: TTI bundling's staged weapon, a separate gambling policy for voice, the most profitable feature per configuration line, the control channel's quiet failures, the power-control floor, HARQ as a voice feature, and the dignified exit — with six KPI-optimization case studies.
TTI Bundling (FAJ 121 2051 / CXC4011253) sends one voice transport block in four consecutive subframes as a single HARQ unit — roughly +4 dB of uplink link budget bought with PUSCH time, for exactly the UE that needs it: power-limited, at the edge, holding a call. The node holds the most instructive configuration posture of the entire kget — the weapon is staged in three positions: trigger thresholds tuned on all nine cells, arming on handover and re-establishment configured (a bundling-eligible UE arriving by mobility is evaluated immediately), and the feature DEACTIVATED — while the golden file says ACTIVATED.
This is the only first-order golden deviation on the node, and it is a governance question, not a config error: someone tuned everything, then switched it off (or never on). Module 8's audit classifies it; Module 9 ranks resolving it first — classify, find the history, then trial on edge-concentrated cells where pmPdcpPktLostUlQci[1] justifies it. Interaction warnings from the FD: frequency hopping and bundling contend for resource patterns; service-specific HARQ multiplies with bundling's repetitions.
Link adaptation picks MCS against a BLER working point — a gambling policy. Data tolerates 10% first-transmission loss because HARQ recovers cheaply within TCP's patience. Voice cannot wait for deep ladders. Robust Uplink Link Adaptation for VoLTE (FAJ 121 5699) gives QCI-1 its own target: ulHarqVolteBlerTarget=5 on this node — half the data working point. The scheduler picks one notch more conservative MCS for voice, spending spectral efficiency (which voice barely uses) to keep first transmissions landing.
The verification surface is Lesson 4-6's leading indicator: mean HARQ transmissions per voice TB. The target halving should hold the mean near 1.05–1.15 even at edge; a climb toward 2 says the channel is outrunning the policy and the deeper tools (bundling, power) are due.
A 30-byte AMR frame travels under ~40 bytes of IPv4/UDP/RTP headers (60 for IPv6) — the headers outweigh the payload. Robust Header Compression (FAJ 121 0892, ACTIVATED) compresses the stack to 1–3 bytes steady-state (RFC 3095 profile 0x0001 RTP/UDP/IP), nearly halving the over-the-air voice payload. Halved payload is not just capacity: smaller TBs survive worse SINR at the same power, so ROHC is also a coverage feature — the cheapest +dB on the node. The state machine (IR → FO → SO) resynchronizes on loss; the practical audit is simply that compression engagement counters move and PDCP-layer voice throughput per call sits near the compressed expectation.
Voice fails on PDCCH differently than on PUSCH: a lost DCI means the UE never knew a grant existed — nothing transmits, nothing HARQs, the loss is silent and multiplicative (control failures gate every dependent traffic success). The node runs the PDCCH outer loop at −70 / 6 / 20 (initial offset, step up, step down in tenths of dB) — an asymmetric loop that backs off aggressively on NACK evidence and creeps back slowly. PDCCH Power Boost (FAJ 121 3057) is operable machinery armed at zero: pdcchPowerBoostMax=0 — the magazine is loaded, the charge is empty, golden-compliant. Coverage Extension (FAJ 121 4629) — the deeper rural tool (PDCCH repetition) — is absent from this dense grid, correctly.
Healthy CQI + voice loss + DCI miss statistics = control-channel problem (Lesson 7-4's paradox). The first dial is the boost magazine: raise pdcchPowerBoostMax on evidence of aggregation-level saturation at cell edge, and watch the cost — boost steals from the PDSCH power budget.
Uplink power control sets the noise floor everyone shares. The node's posture: pZeroNominalPusch=−103 dBm, pZeroNominalPucch=−117 dBm, alpha=1 (full pathloss compensation). Full compensation means edge UEs transmit at whatever it takes to hit the target — friendly to the serving cell's voice, expensive in inter-cell interference. The PUCCH floor 14 dB below PUSCH reflects PUCCH's format robustness. These three numbers are the substrate every Module 4 feature stands on: bundling's +4 dB, the BLER target's conservatism, and the boost magazine all assume this floor. Touch them last, market-wide, never per-complaint.
Service-Specific HARQ (FAJ 121 4645 / CXC4012070, ACTIVATED, enableServiceSpecificHARQ=true ×9) lets voice run its own maximum HARQ transmission count, decoupled from data. The arithmetic is the lesson: each HARQ round trip is 8 ms; a 7-rung ladder spends 56 ms of the 80 ms PDB — survivable exactly once per frame. RLC UM's tReorderingUl/Dl=60 ms (Module 7) is sized to outlast that ladder, barely. Mean HARQ transmissions per voice TB is the network's best leading indicator: it moves days before loss counters, because the ladder absorbs degradation until it can't.
Every call at the true edge ends somewhere. The out-of-coverage machinery (Lesson 4-7) decides whether it ends as an escape (SRVCC retreat to UTRAN — Module 6), a managed release (RLF declared honestly, context cleaned, UE freed to reselect), or a silent hang (the failure mode all of this exists to prevent: a zombie call neither side can end). The boundary features' verdicts land in the release taxonomy — abnormal-release counters at the edge are partially policy, the dignified ending of calls that outran physics. Audit them against geography before treating them as faults.
The loop, as always: symptom → evidence → cause → exact action → verification → rollback.
pmPdcpPktLostUlQci[1]) concentrates on edge-heavy cells; golden audit flags CXC4011253 DEACTIVATED vs golden ACTIVATED.ulHarqVolteBlerTarget 5→10 to match data; two weeks later voice MOS complaints rise with no loss-counter change at first glance.pdcchPowerBoostMax from 0 in one agreed step on the two cells (FAJ 121 3057 is already operable — the magazine is staged for exactly this).