VoIP Quality · Version Comparison · PK_PK
4.11.3 — Caller × Callee Quality Matrix
2026-06-25  ·  19,751 matched call pairs  ·  caller_pp = PK_PK  ·  roomid_match_7d_di
Four metrics · 2×2 heatmap · delta vs baseline D (other/other)
caller side
First Audio Frame ↓ lower
callee: other
callee: 4.11.3
caller
other
D base 658.9 ms
C −183 475.7 ms
caller
4.11.3
B −29 629.8 ms
A −180 478.9 ms
Callee controls. C ≈ A and B ≈ D — columns are nearly identical; rows are nearly identical. Callee 4.11.3 saves −183ms; caller 4.11.3 saves only −29ms (6× asymmetry).
callee side
Init Connection ↓ lower
callee: other
callee: 4.11.3
caller
other
D base 1357 ms
C +187 1544 ms
caller
4.11.3
B +132 1489 ms
A +242 1598 ms
Both roles regress independently. 4.11.3 adds latency on the callee side (+187ms) and the caller side (+132ms). The gradient runs D→C→B→A — no interaction; effects stack.
caller side · video calls only
Audio Quality (MOS) ↑ higher
callee: other
callee: 4.11.3
caller
other
D base 3.407
C −0.037 3.370
caller
4.11.3
B −0.036 3.371
A +0.112 3.519
Cooperation effect. B and C (single-side upgrade) sit below baseline. Only A — both sides on 4.11.3 — gains +0.11 MOS. Consistent with a mutually-negotiated video codec path.
caller side
Send Packet Loss ↓ lower
callee: other
callee: 4.11.3
caller
other
D base 12.35 %
C +1.55 13.90 %
caller
4.11.3
B +0.28 12.63 %
A +0.43 12.78 %
Group C is the outlier (+1.55pp) — no clear version-driven pattern otherwise. May reflect cohort composition differences (network quality of the other/4.11.3 caller group).
Pattern summary
Pattern A · first audio
6×
Callee effect vs caller effect
Callee 4.11.3 → −183ms. Caller 4.11.3 → −29ms. Upgrade the callee, not the caller, to move first-audio latency.
Pattern B · init connection
+18%
Both roles regress with 4.11.3
Callee role adds +187ms; caller role adds +132ms. Effects are additive. 4.11.3 improves first-audio but costs init conn time.
Pattern C · audio quality · video only
∩
Mutual upgrade required for MOS gain
Single-side upgrade leaves quality flat or below baseline. Only when both caller and callee are on 4.11.3 does MOS rise (+0.11).
Raw data — all columns
caller callee calls first_audio init_conn send_loss recv_loss audio_score first_audio init_conn send_loss recv_loss audio_score ring→accept
← caller side ← callee side
4.11.3 4.11.3 2,402 478.9 ms 333.0 ms 12.78 % 4.75 % 3.519 870.3 ms 1,598.4 ms 12.80 % 4.55 % 3.421 7.4
4.11.3 other 6,322 629.8 ms 298.1 ms 12.63 % 4.38 % 3.371 951.0 ms 1,488.8 ms 11.91 % 3.97 % 3.385 7.1
other 4.11.3 3,202 475.7 ms 267.9 ms 13.90 % 4.65 % 3.370 929.4 ms 1,543.7 ms 13.34 % 4.20 % 3.254 6.7
other other 7,825 658.9 ms 248.1 ms 12.35 % 4.41 % 3.407 1,042.0 ms 1,356.8 ms 12.39 % 4.09 % 3.361 6.6
c.* = caller side  ·  e.* = callee side  ·  ring→accept unit unclear (values 6–7, column name says ms but likely seconds)  ·  RTT excluded — string field, AVG corrupted (range: −3,941ms to +470,632ms)  ·  not_ring_pct = 0.0 across all groups (callee_is_ringing encoding unclear)  ·  MOS (audio_score) applies to video calls only  ·  Baseline row D (other/other) shown at reduced opacity