The best mobile networks in Europe now deliver calls where 90% of test calls score above 4.7 on the 5-point MOS scale, according to the 2026 umlaut and connect Mobile Network Tests in the Netherlands, Germany, Austria and Switzerland. The same methodology exposes how uneven that quality is: on UK trains, call success ratios fell to between 87.7% and 92.2%, and over-the-top app calls dropped to a P10 MOS of 3.2. Voice still carries 27% of the score in these benchmarks, and HD voice (VoLTE or Wi-Fi calling) already handles between 93.7% and 99.3% of crowdsourced calls depending on the operator. For the network rollout behind those numbers, see our VoLTE and VoNR deployment statistics. We aggregated data from umlaut (part of Accenture) and connect, the Telecom Regulatory Authority of India, ITU-T, 3GPP, Fraunhofer IIS, Ericsson, Ookla RootMetrics and other primary sources listed in the methodology.
TL;DR
- Best-in-class VoLTE speech quality: P10 MOS-LQO of 4.7 in Dutch, German, Austrian and Swiss city tests (umlaut and connect Mobile Network Tests 2026).
- Voice telephony is worth 27% of the 1,000-point umlaut and connect score, data 48% and crowdsourcing 25% (umlaut and connect methodology 2026).
- KPN and Odido scored 269 of 270 voice points in the Netherlands; EE led the UK on 255 and Three trailed on 221 (umlaut and connect 2026).
- UK call success on trains: 92.2% (EE), 90.5% (Vodafone), 87.9% (Three), 87.7% (VMO2) (umlaut and connect UK 2026).
- Dutch call success on trains: 100% (KPN), 99.6% (Vodafone), 99.3% (Odido) (umlaut and connect Netherlands 2026).
- Fastest call setup: 0.5 seconds at P90 for KPN; slowest: 3.2 seconds for Vodafone UK on roads and trains (umlaut and connect 2026).
- Crowdsourced HD voice share: 99.3% for KPN and Odido, 93.7% for VMO2 (umlaut and connect 2026).
- App-based calls (Opus over data) scored a P10 of 3.5 to 3.9 in UK city drive tests, versus 4.4 to 4.5 for VoLTE (umlaut and connect UK 2026).
- In Chennai, TRAI recorded 0 dropped calls out of 500 on Airtel and 12 out of 499 on BSNL (TRAI Independent Drive Test, July 2026).
- 7.8 billion subscriptions with IMS-based 4G and 5G voice forecast by end of 2029 (Ericsson Mobility Report, voice outlook).
- EVS runs from 5.9 to 128 kbit/s and is supported by more than 200 device models (Fraunhofer IIS).
- POLQA (ITU-T P.863) is in its third edition, approved in March 2018 (ITU-T).
1. How Benchmarks Weigh Voice: 27% of the Score, and Where Operators Land
Voice calls are a shrinking share of what people do with a phone, but they still decide benchmark results. In the umlaut and connect methodology used across Europe, voice connections account for 27% of the 1,000-point total, against 48% for data and 25% for crowdsourcing (umlaut and connect, Mobile Network Test Netherlands 2026). The DACH report says it plainly: voice telephony ‘is still important and should work reliably when you need it’, and connect’s German methodology page confirms the same 27% weighting.
The voice scores show a continent that has largely solved telephony, with one clear exception. Dutch and Swiss operators sit within two points of the 270-point maximum. Germany is close behind, with Telekom on 268. The UK spreads from 255 (EE) to 221 (Three), a 34-point gap that is larger than the gap between the best and worst operators in the Netherlands across the entire 1,000-point test (988 vs 972).
| Metric | Value | Source |
|---|---|---|
| Voice weight in overall score | 27% (data 48%, crowdsourcing 25%) | umlaut and connect methodology 2026 |
| Netherlands voice score (max 270) | KPN 269, Odido 269, Vodafone 268 | umlaut and connect Netherlands 2026 |
| Germany voice score | Telekom 268, Vodafone 263, Telefonica 258 | umlaut and connect DACH 2026 |
| Austria voice score | Magenta 267, A1 263, Drei 250 | umlaut and connect DACH 2026 |
| Switzerland voice score | Swisscom 268, Sunrise 268, Salt 268 | umlaut and connect DACH 2026 |
| UK voice score | EE 255, Vodafone 234, VMO2 229, Three 221 | umlaut and connect UK 2026 |
| Voice samples collected | UK 29,611; Germany 25,446; Switzerland 16,758; Netherlands 16,122; Austria 13,853 | umlaut and connect 2026 reports |
| Drive test distance | UK 10,170 km; Germany 11,010 km; Netherlands 7,130 km | umlaut and connect 2026 reports |
Context note: the DACH tests ran in October 2025 (published in connect issue 1/2026), the UK test from October 27 to November 8, 2025, and the Dutch test from January 14 to February 3, 2026. All are branded 2026 tests. The full UK and DACH results are in the UK 2026 report and the DACH 2026 report.
2. Speech Quality: MOS-LQO Scores Cluster Between 4.0 and 4.7
MOS (mean opinion score) runs from 1 (bad) to 5 (excellent). Benchmarkers report the P10 value, the score that 90% of calls beat, which is a stricter view than an average because it captures the bad calls. umlaut measured it with the POLQA wideband method on Samsung Galaxy S24 Ultra phones while data traffic ran in the background. The headline is that a clean VoLTE call in a European city now tops out at 4.7 MOS-LQO, the highest P10 value recorded in any of the 2026 reports (umlaut and connect, DACH 2026).
The useful signal is in the spread. Within cities, almost every operator lands between 4.4 and 4.7, so the codec and network core are rarely the problem. Quality drops when the radio link gets hard: on UK trains the P10 fell to 4.0 for VMO2 and 4.1 for Vodafone, and on German trains to 4.3 for Telefonica and Vodafone. When the best networks cluster within 0.3 points of each other, a 0.5 to 0.7 point drop is a large gap, and it shows up exactly where people take calls on the move. For how the same scale is used to grade synthetic voices, see our synthetic speech naturalness MOS statistics.
| Metric | Value | Source |
|---|---|---|
| Netherlands cities walk test, P10 MOS-LQO | 4.7 for KPN, Odido and Vodafone | umlaut and connect Netherlands 2026 |
| Germany cities drive test, P10 MOS-LQO | Telekom 4.7, Vodafone 4.6, Telefonica 4.5 | umlaut and connect DACH 2026 |
| Austria cities drive test, P10 MOS-LQO | Magenta 4.7, A1 4.6, Drei 4.5 | umlaut and connect DACH 2026 |
| Switzerland cities drive test, P10 MOS-LQO | Sunrise 4.7, Salt 4.7, Swisscom 4.6 | umlaut and connect DACH 2026 |
| UK cities drive test, P10 MOS-LQO | EE 4.5; Vodafone, Three and VMO2 4.4 | umlaut and connect UK 2026 |
| UK roads drive test, P10 MOS-LQO | EE 4.3, Vodafone 4.2, VMO2 4.2, Three 4.1 | umlaut and connect UK 2026 |
| UK railways, P10 MOS-LQO | EE 4.2, Three 4.2, Vodafone 4.1, VMO2 4.0 | umlaut and connect UK 2026 |
| Germany railways, P10 MOS-LQO | Telekom 4.4, Telefonica 4.3, Vodafone 4.3 | umlaut and connect DACH 2026 |
Context note: MOS-LQO values from different POLQA modes (narrowband vs wideband vs super-wideband) are not directly comparable, and umlaut reports wideband. A P10 of 4.7 does not mean every call sounded perfect; it means 10% of calls scored below 4.7.
3. Call Reliability: Near-Perfect in Cities, Fragile on Trains
Success ratio (calls that connect and complete without dropping) is where the real differences between networks hide. In cities, every operator tested in the Netherlands, Germany, Austria and Switzerland completed at least 99.2% of calls. On UK trains, success fell to 87.7% for VMO2 and 87.9% for Three, so roughly one call in eight failed on those networks (100 minus 87.7 = 12.3%). umlaut’s own commentary calls voice on UK railways ‘clearly an area where all four UK operators show some need for improvements’.
The contrast with the Netherlands, where KPN completed 100% of calls on trains, shows this is a network investment issue rather than a physics problem. Germany sits in between: Telekom reached 98.4% on trains and Telefonica 96.1%, and connect credits part of the improvement to Deutsche Bahn fitting trains with RF-transparent windows. Outside Europe, regulator drive tests show the same pattern of a clear laggard: in TRAI’s July 2026 Chennai test, Airtel dropped 0 of 500 calls while BSNL dropped 12 of 499 (2.4%) (TRAI Press Release 118/2026).
| Metric | Value | Source |
|---|---|---|
| UK railways call success ratio | EE 92.2%, Vodafone 90.5%, Three 87.9%, VMO2 87.7% | umlaut and connect UK 2026 |
| UK roads call success ratio | EE 98.2%, Vodafone 97.6%, VMO2 97.1%, Three 95.4% | umlaut and connect UK 2026 |
| Netherlands railways call success ratio | KPN 100%, Vodafone 99.6%, Odido 99.3% | umlaut and connect Netherlands 2026 |
| Germany railways call success ratio | Telekom 98.4%, Vodafone 97.9%, Telefonica 96.1% | umlaut and connect DACH 2026 |
| Austria railways call success ratio | Magenta 96.1%, A1 95.9%, Drei 95.8% | umlaut and connect DACH 2026 |
| Switzerland railways call success ratio | Swisscom 99.8%, Sunrise 99.2%, Salt 98.6% | umlaut and connect DACH 2026 |
| Chennai dropped calls (calls established) | Airtel 0/500, Jio 1/514, Vi 1/508, BSNL 12/499 | TRAI Independent Drive Test, July 2026 |
| Chennai calls with silence over 3 seconds | BSNL 31 of 486 calls (38 instances); Airtel 3/493; Jio 2/492; Vi 3/501 | TRAI Independent Drive Test, July 2026 |
Context note: TRAI’s Chennai test covered 232.8 km of city driving, 7 hotspots and 9.9 km of walk tests from July 15 to 17, 2026, so each operator’s figures rest on about 500 calls. TRAI lists speech quality (MOS) among its assessed parameters but did not publish MOS values in this press release. Over a whole call, a mute of more than 3 seconds hurts perceived quality as much as a low MOS, which is why regulators now count silences separately.
4. Call Setup Time: From Half a Second to Over Three
Call setup time is the gap between pressing call and hearing it ring, and it is the metric where VoLTE configuration choices show most clearly. umlaut reports it as a P90 percentile value rather than an average, so it reflects the slower calls in each sample, not the best case. KPN connected calls in 0.5 seconds at P90 in every Dutch city scenario, while Vodafone UK needed 2.7 seconds in cities and 3.2 seconds on roads and trains.
The setup time ranking does not follow speech quality. Vodafone UK matched EE’s 4.6 P10 MOS in city walk tests but took almost twice as long to set up calls (2.6 vs 1.4 seconds). In Austria, Drei, the only operator measured in 5G standalone mode, posted setup times of 1.6 to 1.9 seconds; connect notes that fallback from 5G SA to VoLTE ‘poses additional technical challenges’, which also cost Drei some success ratio. This is the practical cost of running 5G SA before voice over 5G (VoNR) is in place, a transition covered in our 5G adoption statistics.
| Metric | Value | Source |
|---|---|---|
| Netherlands cities, P90 call setup time | KPN 0.5 s, Odido 0.6 s, Vodafone 1.0 s | umlaut and connect Netherlands 2026 |
| Switzerland cities, P90 call setup time | Sunrise 0.8 s, Salt 0.9 s, Swisscom 1.1 s | umlaut and connect DACH 2026 |
| Germany cities, P90 call setup time | Telekom 1.0 s, Telefonica 1.1 s, Vodafone 1.1 s | umlaut and connect DACH 2026 |
| Austria cities drive test, P90 call setup time | Magenta 1.0 s, A1 1.2 s, Drei 1.6 s | umlaut and connect DACH 2026 |
| Austria roads, Drei P90 call setup time (5G SA mode) | 1.9 s | umlaut and connect DACH 2026 |
| UK cities drive test, P90 call setup time | EE 1.5 s, VMO2 1.6 s, Three 1.8 s, Vodafone 2.7 s | umlaut and connect UK 2026 |
| UK roads and railways, Vodafone P90 call setup time | 3.2 s | umlaut and connect UK 2026 |
| Germany railways, P90 call setup time | Telekom 1.1 s, Telefonica 1.3 s, Vodafone 1.3 s | umlaut and connect DACH 2026 |
Context note: setup time here is measured mobile to mobile within the same network; calls to other networks or to landlines can take longer. All operators except Drei in Austria carried test calls exclusively over VoLTE, per the connect methodology.
5. HD Voice Adoption: 93.7% to 99.3% of Real-World Calls
Drive tests show what a network can do; crowdsourcing shows what users actually get. umlaut’s HD voice metric is the share of users’ voice connections set up over VoLTE or Wi-Fi calling, and therefore capable of HD quality. It is drawn from very large panels: about 12.4 billion samples from almost 2.6 million users in Germany, 5,702 million samples in the UK and 1,740 million in the Netherlands, according to the 2026 reports.
The figures confirm that HD voice is now the default in Western Europe, but not universal. KPN and Odido reach 99.3%, while VMO2 sits at 93.7% and Vodafone UK at 94.6%, meaning roughly one call in 16 on VMO2 is still set up without VoLTE or Wi-Fi calling (100 minus 93.7 = 6.3%). Globally, Ericsson’s voice and communication services outlook expects VoLTE subscriptions of 6.3 billion at the end of 2024 and 7.8 billion subscriptions with IMS-based 4G and 5G voice by the end of 2029, about 90% of all 4G and 5G subscriptions.
| Metric | Value | Source |
|---|---|---|
| HD voice share, Netherlands | KPN 99.3%, Odido 99.3%, Vodafone 98.8% | umlaut and connect Netherlands 2026 (crowdsourcing) |
| HD voice share, Germany | Telekom 99.1%, Telefonica 98.0%, Vodafone 96.6% | umlaut and connect DACH 2026 (crowdsourcing) |
| HD voice share, Austria | Magenta 98.7%, Drei 98.5%, A1 97.5% | umlaut and connect DACH 2026 (crowdsourcing) |
| HD voice share, Switzerland | Swisscom 98.3%, Salt 98.3%, Sunrise 98.2% | umlaut and connect DACH 2026 (crowdsourcing) |
| HD voice share, UK | EE 99.1%, Three 98.7%, Vodafone 94.6%, VMO2 93.7% | umlaut and connect UK 2026 (crowdsourcing) |
| Crowdsourced samples analysed | Germany 12.4 billion; UK 5,702 million; Netherlands 1,740 million | umlaut and connect 2026 reports |
| VoLTE subscriptions, end of 2024 | 6.3 billion | Ericsson Mobility Report, voice outlook |
| IMS voice subscriptions forecast, end of 2029 | 7.8 billion (about 90% of 4G and 5G) | Ericsson Mobility Report, voice outlook |
Context note: HD voice share measures the bearer (VoLTE or VoWiFi), not the codec or the score of each call. A VoLTE call can still run on narrowband AMR if one handset or the interconnect does not support wideband.
6. Carrier Calls vs App Calls: The OTT Quality Gap
Most people now place a share of their calls through messaging apps, and the 2026 benchmarks measure that too. umlaut’s ‘conversational app’ test sets up a voice channel over mobile data using SIP, STUN and the Opus codec, the same building blocks used by popular calling apps. The result is consistent across every country: app calls score lower than operator VoLTE calls on the same routes. In UK city drive tests the app P10 was 3.5 to 3.9, against 4.4 to 4.5 for VoLTE (umlaut and connect, UK 2026).
The gap widens as conditions get worse, because app audio competes with everything else on the data connection while VoLTE gets a dedicated quality-of-service bearer. On UK trains the app P10 fell to 3.2 to 3.4, and on Austrian roads Drei’s app calls dropped to 3.4 with a 97.7% success ratio. Dutch networks are the exception that proves the point: with abundant 5G capacity, app calls held 4.1 to 4.4 in city drive tests. Our VoIP and UCaaS statistics cover the business side of the same shift.
| Metric | Value | Source |
|---|---|---|
| UK cities drive test, app call P10 MOS-LQO | EE 3.9, Vodafone 3.6, VMO2 3.6, Three 3.5 | umlaut and connect UK 2026 |
| UK railways, app call success / P10 MOS-LQO | EE 98.1% / 3.4; Vodafone 96.8% / 3.3; Three 96.2% / 3.3; VMO2 95.1% / 3.2 | umlaut and connect UK 2026 |
| Netherlands cities drive test, app call P10 MOS-LQO | KPN 4.4, Odido 4.3, Vodafone 4.1 (100% success) | umlaut and connect Netherlands 2026 |
| Germany cities drive test, app call P10 MOS-LQO | Telekom 4.4, Telefonica 4.1, Vodafone 4.0 | umlaut and connect DACH 2026 |
| Germany railways, app call P10 MOS-LQO | Telefonica 3.8, Telekom 3.7, Vodafone 3.4 | umlaut and connect DACH 2026 |
| Switzerland cities drive test, app call P10 MOS-LQO | 4.3 for Swisscom, Sunrise and Salt | umlaut and connect DACH 2026 |
| Share of samples meeting the OTT voice latency class (100 ms round trip) | UK 93.6% to 95.2%; Netherlands 98.8% to 99.2% | umlaut and connect 2026 (crowdsourcing) |
Context note: the conversational app test runs on a separate phone over the data connection, so it measures the network’s data path rather than any specific app. App call quality also depends on the app’s own codec settings and servers, which a network benchmark cannot control.
7. The Standards Behind the Numbers: POLQA, MOS and EVS
Every MOS figure above traces back to a short list of ITU-T and 3GPP documents. The subjective method, people listening and rating calls, is still defined by ITU-T P.800, in force since August 1996 (ITU-T P.800), with MOS terminology (MOS-LQO for objective listening-quality scores) set out in P.800.1, July 2016 (ITU-T P.800.1). The objective model that benchmarkers actually run is POLQA, ITU-T P.863, whose current third edition was approved in March 2018 with an amendment in April 2020 (ITU-T P.863). For network planning rather than measurement, operators use the E-model, ITU-T G.107 (June 2015 edition), which outputs an R-factor (ITU-T G.107).
On the codec side, the step change is from AMR-WB to EVS. AMR-WB is the original HD voice codec. EVS, specified in 3GPP TS 26.441 for Release 12, had its first version approved on June 16, 2014 and reached version 19.0.0 on October 22, 2025 (3GPP specification portal). Fraunhofer IIS says EVS runs from 5.9 to 128 kbit/s, delivers super-wideband audio from 9.6 kbit/s and is supported by more than 200 device models (Fraunhofer IIS). For the human side of what that bandwidth means, our phone call statistics track how often people still pick up.
| Metric | Value | Source |
|---|---|---|
| Subjective transmission quality method | ITU-T P.800, in force since August 1996 | ITU-T |
| MOS terminology standard | ITU-T P.800.1, July 2016 edition | ITU-T |
| EVS specification (TS 26.441) | v1.0.0 approved June 16, 2014 (Release 12); v19.0.0 on October 22, 2025 | 3GPP specification portal |
| EVS bit rates and device support | 5.9 to 128 kbit/s; super-wideband from 9.6 kbit/s; more than 200 device models | Fraunhofer IIS |
| E-model for transmission planning | ITU-T G.107, June 2015 edition | ITU-T |
Context note: The P.800 baseline dates from 1996 and the G.107 edition from 2015; these are the most recent in-force editions, not stale data.
Summary: Cellular Voice Quality (MOS) by the Numbers
| Metric | Value | Source |
|---|---|---|
| Best P10 speech quality, VoLTE | 4.7 MOS-LQO (NL, DE, AT, CH city tests) | umlaut and connect 2026 |
| UK cities drive test P10 speech quality | 4.4 to 4.5 MOS-LQO | umlaut and connect UK 2026 |
| Lowest P10 speech quality measured (UK railways) | 4.0 (VMO2) | umlaut and connect UK 2026 |
| Voice weight in benchmark score | 27% | umlaut and connect methodology 2026 |
| Netherlands voice score | 269 / 269 / 268 of 270 | umlaut and connect Netherlands 2026 |
| Switzerland voice score | 268 for all three operators | umlaut and connect DACH 2026 |
| UK voice score range | 255 (EE) to 221 (Three) | umlaut and connect UK 2026 |
| UK railways call success | 87.7% to 92.2% | umlaut and connect UK 2026 |
| Netherlands railways call success | 99.3% to 100% | umlaut and connect Netherlands 2026 |
| Fastest P90 call setup | 0.5 s (KPN) | umlaut and connect Netherlands 2026 |
| Slowest P90 call setup | 3.2 s (Vodafone UK, roads and trains) | umlaut and connect UK 2026 |
| Drei (5G SA mode) P90 call setup, cities | 1.6 s | umlaut and connect DACH 2026 |
| Highest crowdsourced HD voice share | 99.3% (KPN, Odido) | umlaut and connect Netherlands 2026 |
| Lowest crowdsourced HD voice share | 93.7% (VMO2) | umlaut and connect UK 2026 |
| App call P10, UK cities drive test | 3.5 to 3.9 | umlaut and connect UK 2026 |
| App call P10, UK railways | 3.2 to 3.4 | umlaut and connect UK 2026 |
| Chennai dropped calls | Airtel 0/500; BSNL 12/499 | TRAI, July 2026 |
| IMS voice subscriptions, 2029 forecast | 7.8 billion | Ericsson Mobility Report |
| EVS bit-rate range | 5.9 to 128 kbit/s | Fraunhofer IIS |
| Current POLQA edition | ITU-T P.863 (03/18) | ITU-T |
Methodology and Sources
Every figure above was read during research for this article and traced to the organization that produced it. The umlaut and connect results come from the full published PDF reports, not press summaries.
- umlaut (part of Accenture) and connect: Mobile Network Test Netherlands 2026, Mobile Network Test UK 2026, Great Mobile Network Test 2026, Germany, Austria and Switzerland, connect test methodology 2026 (German), connect testlab overview
- Deutsche Telekom: press release on the connect 2026 test (operator statement, used only for context)
- Telecom Regulatory Authority of India: Chennai Independent Drive Test press release, PR No. 118/2026 (full PDF)
- ITU-T: P.800, P.800.1, P.863 (POLQA), G.107 (E-model)
- 3GPP: TS 26.441, Codec for Enhanced Voice Services, general overview
- Fraunhofer IIS: EVS codec page
- Ericsson: Mobility Report, voice and communication services outlook
- Ookla RootMetrics: US State of the Mobile Union 1H 2026, as reported by Light Reading (243,000+ drive-miles, 6,900+ indoor tests, about 3 million tests from January to June 2026)
- Data watch: Public MOS data is concentrated in Europe because umlaut and connect publish full per-operator voice tables; the US RootMetrics 1H 2026 release covered speed, reliability and 5G but its public summary contained no call failure or speech quality figures, so no US voice metrics are cited. Opensignal’s 2026 Voice App Experience reports and Ofcom’s Mobile Matters 2025 report blocked automated access (HTTP 403) and were excluded rather than quoted from search snippets. TRAI lists MOS among its drive-test parameters but publishes only dropped calls and call silences in its press releases. The UK and DACH campaigns were measured in October and November 2025 and the Dutch campaign in January and February 2026; all are labelled 2026 by the publishers. umlaut’s MOS figures use POLQA wideband and are P10 values, so they should not be compared with average MOS figures or super-wideband scores from other sources. Ericsson’s 6.3 billion VoLTE figure is an end-2024 estimate. ITU-T P.800 (1996) and G.107 (2015) are older baselines that remain the in-force editions (most recent available data).
Last updated: October 3, 2026. We update this roundup quarterly, and the next refresh is expected when umlaut and connect publish their next Mobile Network Tests for Germany, Austria, Switzerland and the UK (connect issue 1/2027) and TRAI releases further Independent Drive Test reports.