Voice quality end to end: the codec-rate loop the RAN can close, RLC's two worldviews, the 280-millisecond corridor and its queue defenders, packet-loss forensics, and the transport boundary where DSCP promises are proven — with five KPI-optimization case studies.
The phones negotiate the codec (Module 1's SDP); the network knows the radio. Classically the codec adapts blindly — after damage. VoLTE Rate Recommendation (FAJ 121 5014) closes the loop: the eNodeB sends in-band bit-rate recommendations (TS 36.321 MAC CE) and AMR-WB/EVS shift gears mid-call, per direction — down before the cell edge bites, up when headroom returns. The inversion worth teaching: a 23.85 kbps stream losing 8% sounds far worse than a 12.65 kbps stream losing nothing; downshifting is a quality feature, and the lower rate also survives deeper fades — retainability engineering disguised as audio engineering. On this node: bitRateRecommendationEnabled=false on the qci1 profile — staged dark, license-gated per the MOM, reopened by edge-concentrated loss evidence.
AM spends time to guarantee delivery (status reports, polls, retransmission); UM spends packets to guarantee time (numbering, a window, and nothing else). Voice can only choose UM: a frame delivered late by an RLC retry cycle is a dead frame the de-jitter buffer discards — integrity is timely delivery. SIP rides AM, where persistence is right. The numbers are cut to the deadline: voice runs 10-bit RLC / 12-bit PDCP sequence numbers (small, frequent packets; headers matter at 30-byte payloads — Module 4's ROHC arithmetic), data runs 16/15 for CA-scale flight volumes, and tReorderingUl/Dl=60 ms is sized to outlast the 7-rung HARQ ladder (7×8 ms) — barely, deliberately. The AM refinements on the SIP bearer (Adaptive RLC Poll-Retransmission, Load-Based UL RLC Retx Threshold) tune persistence's enthusiasm without abandoning it.
The mode someone “improved”: QCI-1 set to AM made the loss KPI fall and complaints explode — frames arrived embalmed, 60–100 ms late, and the de-jitter buffer stretched then dropped them. Loss improved + jitter exploded = wrong ruler. The review checklist owns this rule now.
Mouth-to-ear at ~280 ms conversation begins to die (ITU-T G.114's working region). The corridor is shared: codec framing ×2 (~20 ms + encode), the de-jitter buffer's deliberate hold, core + transport transit (possibly ×2 operators), and two radio legs inheriting the remainder — which is why Module 1 gave the eNodeB pdb=80. Jitter is the hidden tax: low average + high variance forces deep buffers, and a packet beyond the buffer's patience is functionally lost however alive. The defender is per-service AQM, configured on this node exactly as the physics demand: aqmMode=MODE2 on qci1 with pdbOffset=50 (deadline-respecting voice variant), OFF on qci5 (never drop SIP to manage latency — AM retransmits anyway), MODE1 on qci9 (classic early-drop; TCP understands loss). Three relationships with time, live in the kget.
The diagnostic discipline is subtraction: radio variance from HARQ/scheduler statistics, transport from TWAMP, the residual attributed honestly to the far side. The bufferbloat case is the genre: talk-over at 300 ms, radio slices normal, one new router buffering deeply mid-corridor — no loss, beautiful throughput, +80 ms queue under load. Voice quality is a supply chain; any unmanaged queue anywhere taxes every call.
The first forensic question is where, because why has different suspects per surface: pmPdcpPktLostUlQci[1] (uplink air — power, adaptation, interference), pmPdcpPktDiscDlPelrQci[1] (in-node discard — queues, AQM, congestion with timestamps), pmPdcpPktDiscDlPelrUuQci[1] (downlink air — transmitted, never acknowledged). Traces convert ratios into stories: loss + collapsing CQI = coverage/interference; loss + healthy CQI + DCI misses = the control channel (Module 4's paradox); loss clustered in handover windows = mobility timing (Module 6). And the ear hears patterns, not percentages: concealment absorbs isolated loss, three consecutive frames is a syllable's corpse — the same 1% can be inaudible scattered or infuriating clustered. Report surface, rate, burstiness, and a labeled perceptual band; let MOS frameworks rank, never replace, the evidence.
The radio's promises end at the cabinet door. The markings as a system: dscp 40 (voice — expedited), 26 (SIP — assured), 10 (data — best-effort intent), with the per-ARP refinement staged unused (dscpArpMap all −1 — emergency could mark differently). The failure mode is silent demotion: one re-marking interface and voice queues behind bulk — invisible until congestion. TWAMP (RFC 5357) is the proof: timestamped probes per marking class, compared under load, per segment. Three rules make it useful: probes must carry the markings (unmarked tests test nothing about voice), windows must be busy-hour (empty networks honor everything), sessions must be per-segment (end-to-end detects, segments diagnose). The maintenance-window case: a vendor swap shipped default queues collapsing all DSCP into one — the radio saw clean air and rising in-node discards; the TWAMP differential told within an hour. The differential is a permanent canary, not a one-time test.
bitRateRecommendationEnabled=false.aqmMode=OFF (this node's posture); keep MODE2+offset50 on qci1, MODE1 on qci9.