4.11.3 — Callee Drives First Audio; Init Conn Regresses on Both Roles

PK·PK corridor · three structural patterns from the 2×2 caller×callee version matrix

Source

roomid_match_7d

Date

2026-06-25

Corridor

PK → PK

Total Calls

19,751

01 Key patterns

Pattern A · first audio

−27%

Callee version alone drives gain

Upgrading the callee to 4.11.3 saves −183ms on caller-side first-audio; upgrading only the caller contributes −29ms. Callee effect is 6× larger.

Pattern B · init connection

+18%

4.11.3 adds latency on both roles

Callee on 4.11.3 adds +187ms to callee init conn; caller on 4.11.3 adds +132ms. Both effects are additive and independent.

Pattern C · audio quality (video calls only)

+0.11

Only mutual upgrade improves score

Single-side upgrades give no improvement (B and C both below baseline D). Only when both sides are 4.11.3 does audio quality rise by +0.11 MOS. MOS applies to video calls only.

02 Pattern A — First audio frame: callee is the controlling variable

First-audio latency (caller-side) groups by callee version, not caller version. When the callee is on 4.11.3, caller first-audio is ~477ms regardless of whether the caller is also on 4.11.3. When the callee is on an older build, caller first-audio is ~644ms regardless of caller version.

Caller First Audio (ms) — regrouped by callee version lower is better

X-axis = caller version; two datasets show callee version. Each pair of bars is nearly identical within a dataset, proving callee version is the sole driver.

Callee version effect

−183 ms

Upgrading callee to 4.11.3 (other→413): avg of (C−D) and (A−B)

Caller version effect

−29 ms

Upgrading caller to 4.11.3 (other→413): avg of (B−D) and (A−C)

03 Pattern B — Init connection: independent regression on both roles

Unlike first-audio, init connection time worsens with 4.11.3 on either side. The two version effects are additive: callee-side 4.11.3 adds ~+187ms; caller-side 4.11.3 adds ~+132ms. Group A (both 4.11.3) is the worst at 1,598ms vs 1,357ms baseline.

Callee Init Connection Time (ms) — all 4 groups lower is better · baseline = D (other/other) = 1,357ms

All three non-baseline groups show elevated init conn. Caller + callee effects add without interaction.

04 Pattern C — Audio quality: cooperation effect

Audio quality score shows a non-linear pattern. Groups B and C (single-side 4.11.3) score below the all-other baseline (D = 3.407). Only group A — both caller and callee on 4.11.3 — delivers a meaningful gain (+0.11 MOS, +3.2%). This suggests a video codec or transport path that activates only when both peers are on the newer build.

Note: MOS is computed for video calls only; the audio quality scores here reflect the video-call subset within PK·PK and are not representative of voice-only calls in this sample.

Caller Audio Quality Score — all 4 groups higher is better · baseline = D (other/other) = 3.407

B and C (single-side upgrade) are below baseline. A (mutual upgrade) is clearly above — a cooperation effect consistent with a mutually-negotiated video codec or transport path. MOS is video-call only; voice-only calls are not captured in this metric.

05 Full metric breakdown

All 8 metrics across the 4 caller×callee version groups. Best cell highlighted teal, worst highlighted red. Delta is vs group D (other/other baseline).

Source: vew.dwd_botim_voip_call_roomid_match_7d_di · filter: caller_start_date = '2026-06-25', caller_pp = 'PK_PK' · 19,751 matched call pairs · Groups: A = 4.11.3/4.11.3 (2,402 calls), B = 4.11.3/other (6,322), C = other/4.11.3 (3,202), D = other/other (7,825) · RTT excluded — string field contains corrupt values (negative/overflow). Not-ring rate = 0.0% across all groups; callee_is_ringing encoding unclear.