DRX for voice: the anatomy of the 40-millisecond contract, the arbitration between services, the adaptive layer, and the economics of the inactivity timer — short by lesson count, dense by consequence, with five KPI-optimization case studies.
No fake things. Every FAJ/CXC, MO attribute, threshold, and feature state below is from the Ericsson CPI (LTE RAN 25.Q4.4), the live NYC node kget (25.Q3), the operator golden file, and the LTE PM counter inventory. 3GPP citations name exact specifications.
CH 1The 40-millisecond contract
DRX is a treaty between battery and latency: the UE may close its receiver, the network promises to transmit only in the agreed windows. The voice profile on this node — DrxProfile=1 — is the voice-shaped rhythm: longDrxCycle=SF40 (40 ms — exactly two voice frames per wake), onDurationTimer=PSF10, drxInactivityTimer=PSF8, drxRetransmissionTimer=PSF2, and shortDrxCycleTimer=0 — one rhythm, no second gear. The arithmetic: a 40 ms cycle with a 10-subframe on-duration keeps the receiver open 25% of the time at idle-talk, while guaranteeing any voice frame waits at most one cycle — inside the 80 ms PDB with room for one full HARQ ladder (Module 4's 56 ms).
Each timer is a clause: onDurationTimer is the listening window; drxInactivityTimer extends wakefulness after activity (8 subframes — short, because voice traffic is metronomic, not bursty); drxRetransmissionTimer keeps the receiver up just long enough for a pending HARQ retransmission (2 subframes of patience).
CH 2Arbitration — the most demanding passenger
A UE has one receiver and possibly many bearers. Service-Specific DRX (FAJ 121 3011, ACTIVATED) maps QCIs to DrxProfiles and arbitrates by drxPriority: the connection follows the most demanding active service. Voice arrives → the voice profile's 40 ms rhythm wins over any data profile's longer sleep; voice ends → the UE relaxes back. The audit point is profile coverage: every QCI in the QciTable must map somewhere deliberate, or new services inherit defaults nobody chose.
CH 3The adaptive layer
Above the static profiles: Adaptive DRX (FAJ 121 4429) adjusts cycle selection from observed traffic, and Efficient DRX/DTX (FAJ 121 0801) is the fleet-economics frame — milliwatts × hours × millions of devices. The voice caution is structural: adaptation logic reading traffic can misread silence — a mid-call mute is not inactivity, it is a conversation breathing (SID frames every 160 ms still flow). Any adaptive layer must treat QCI-1 presence as a hard floor under its decisions, which is exactly why the static voice profile remains the anchor and adaptation is scoped to data.
CH 4Inactivity economics
tInactivityTimer decides when a silent connection is released to idle. Too short: every release-then-redial costs a full setup ladder (Module 2), paging load, and a window of unreachability — and a redial during a dropped-but-not-noticed call looks like a new attempt, polluting accessibility statistics. Too long: connected-state UEs hold PUCCH resources (Module 3's budget) and batteries drain. VoLTE Inactivity Timer Extension (FAJ 121 5181) adds inactivityTimerOffset=35 on this node: a connection that recently carried voice earns 35 extra seconds before release — because post-call signalling (SIP BYE ladders, RTCP, the user redialing the dropped party) clusters right after voice activity ends.
The twin-spike signature
Lesson 5-4's diagnostic, worth engraving: when drops and setup attempts spike together, suspect release policy, not radio — users redialing dropped calls manufacture both numbers at once. Check the inactivity chain before touching any radio parameter.
CH 5Five KPI-optimization case studies
CASE 1
Battery complaints vs the 40 ms contract
KPI symptom
A device OEM reports VoLTE talk-time regression on this market vs a competitor's.
Evidence
Profile dump: competitor runs SF64 voice cycle (riskier vs PDB); this node's SF40/PSF10 keeps 25% receiver duty. UE power traces confirm the delta is DRX duty, not RF.
Root cause
A deliberate latency-vs-battery position, not a fault.
Action
None unilaterally. If revisited: trial onDurationTimer PSF10→PSF8 (duty 20%) before touching the cycle — the cycle is the latency contract, the on-duration is the cheaper dial.
Edge-cell voice loss without CQI degradation after a battery-driven profile edit elsewhere in the region (drxRetransmissionTimer PSF2→PSF1).
Evidence
HARQ retransmissions scheduled while the receiver slept; effective ladder depth halved at edge; pmPdcpPktLostUlQci[1] up only where mean-HARQ was already >1.3.
Root cause
DRX clause broke HARQ's assumption — the timers are one system (Modules 4+5), not two.
Action
Restore PSF2. Add a config-review rule: DRX edits require the HARQ interaction check.
Verification
Edge loss returns to baseline; the rule enters the golden-audit exception logic.
Rollback
The restore.
CASE 3
Accessibility “improvement” that was a release storm
KPI symptom
Setup attempts +18% week over week; management reads it as demand growth.
Evidence
The twin spike: abnormal releases rose in lockstep; per-IMSI traces show redial pairs seconds after releases; a neighboring team had cut tInactivityTimer to harvest PUCCH headroom.
Root cause
Premature releases manufacturing both attempts and drops.
Action
Restore the timer; if PUCCH pressure is real, solve it in Module 3's budget (dimensioning), not by churning connections.
Verification
Both spikes collapse together — the signature confirms the diagnosis.
Post-call signalling histogram from traces: SIP BYE / re-INVITE / redial activity concentrates in the 25 s after voice ends; without the offset these become full setup ladders.
Root cause
n/a — this is the defense of an existing optimization with data.
Action
Keep 35; document the histogram in the golden file's rationale field so the next review reads evidence, not folklore.
Verification
The recurring audit (Module 8) carries the rationale forward.
Rollback
n/a.
CASE 5
The adaptive layer that misread a mute
KPI symptom
Sister market with aggressive adaptive DRX on all QCIs: complaints of clipped first words after silences.
Evidence
Traces: cycles lengthened during mutes (SID-only periods); first talk-burst after silence waits out a long cycle — onset clipping, the latency tail Module 8's dashboard case made visible.
Root cause
Adaptation reading SID silence as inactivity — voice presence must floor the cycle.
Action
Scope FAJ 121 4429 adaptation to data QCIs; QCI-1 pins DrxProfile 1.
Verification
Post-silence onset delay percentiles collapse to one cycle max.
Rollback
Feature scoping attribute.
References
Ericsson CPI, LTE RAN 25.Q4.4 — FAJ 121 3011 (Service-Specific DRX), 4429 (Adaptive DRX), 0801 (Efficient DRX/DTX), 5181 (VoLTE Inactivity Timer Extension)
Live node kget 25.Q3: DrxProfile=1 (SF40/PSF10/PSF8/PSF2, shortDrxCycleTimer=0), inactivityTimerOffset=35