Testing voice agents on UK accents over phone audio

Phone calls to a voice agent usually arrive as 8 kHz audio, which carries less detail than the 16 kHz wide-band audio that published studies say many speech models are trained on. Regional accents add errors on top, and the size of the effect differs from one speech engine to the next.

Key facts

• RFC 3551 lists the phone payload types PCMU and PCMA with a clock rate of 8,000 Hz (IETF RFC 3551).
• One paper on British accents reports a word error rate of 6.0% on clean audio and 12.9% under its telephony condition (Haq et al.).
• Whisper-large-v3 scored 28.7% word error rate on Indian-accented English in 8 kHz customer-support calls (Godambe et al.).
• In the EdAcc paper, the best model scored 19.7% word error rate on EdAcc and 2.7% on clean read US English (Sanabria et al.).
• The OpenSLR 83 page lists 17,877 UK and Ireland English utterances under CC BY-SA 4.0 (OpenSLR 83).

This guide is about understanding: what the phone line does to audio and what the research says. If you want to build and run a test, read how to test a voice agent on accents and background noise. Our own method and data plan are on the UK accents benchmark page, and the product side is on accents and noisy lines.

What does the phone line change in the audio?

One common phone codec is G.711. RFC 3551 lists its two payload types, PCMU and PCMA, with a clock rate of 8,000 Hz (IETF RFC 3551, read 27 October 2026). ITU-T titles the standard "Pulse code modulation (PCM) of voice frequencies" (ITU-T G.711, read 27 October 2026).

A 2014 paper from Technische Universität Braunschweig put it plainly: "most telephone calls are still narrowband", with a speech bandwidth below 4 kHz. It also cited earlier work reporting a relative word error rate (WER) reduction of 20% from doubling the sampling rate from 8 to 16 kHz (Bauer et al., EUSIPCO 2014, read 27 October 2026). The paper takes that figure from a 2001 textbook, so it shows the direction and not what a current model will do.

Your carrier or platform may also transcode a call, so the audio your speech engine receives may differ from what you assume. Test the path your calls take.

What recent studies found

Two 2026 papers measure telephone conditions with recent models.

• A paper on British accents in financial voice assistants tested several speech systems on recordings from 90 speakers. It reports WER of 6.0% on clean audio and 12.9% under its telephony condition, which it describes only as "narrowband codec/bandwidth degradation" (Haq et al., arXiv 2610.06587, 5 October 2026, read 27 October 2026). The paper does not give the codec or sample rate.
• A study of customer-support calls recorded at 8 kHz found Whisper-large-v3 scored 28.7% WER on Indian-accented English. The authors put the high error rates of open-source models "mainly" down to training on "read-style speech in wide-band conditions" (Godambe et al., arXiv 2606.18659, 17 June 2026, read 27 October 2026). That is not a UK accent, but it is real phone audio.

What have studies found on accents?

The Edinburgh International Accents of English Corpus (EdAcc) holds almost 40 hours of video-call conversations between friends. In the paper that introduced it, the best model scored 19.7% WER on EdAcc and 2.7% on clean read US English (Sanabria et al., ICASSP 2023, read 27 October 2026). Those models are now several years old, and EdAcc audio comes from video calls, not phone lines. It measures accent, not bandwidth.

For UK regions specifically, the British accents paper found that effects vary by model. Its Table 1, averaged over six acoustic conditions, shows Qwen3-ASR-1.7B in chunked mode scoring 22.2% WER on Scottish speakers against 12.9% on Southern English speakers. Whisper Turbo scored 16.1% on Scottish speakers against 18.2% on Southern English speakers. So the direction is not fixed, and a result for one engine does not transfer to another (Haq et al., same paper). The same paper found that WER and the agent's tool-call accuracy were strongly linked overall (a correlation of -0.93). The link was much weaker for Scottish speech (-0.29), which the authors put largely down to the Scottish result for the Qwen3 chunked system, so WER alone did not predict whether the agent did the right thing for those speakers.

A study of Newcastle English reports that ASR errors "directly correlate with regional dialectal features" (Serditova, Tang and Steffens, Interspeech 2025, read 27 October 2026).

Public data you can use

• OpenSLR 83 holds UK and Ireland English dialect recordings from 120 volunteers, over 31 hours, covering Southern England, Midlands, Northern England, Welsh, Scottish and Irish varieties (Demirsahin et al., LREC 2020). The OpenSLR page lists 17,877 utterances under CC BY-SA 4.0, and the Irish set it lists is a male archive only (OpenSLR 83, read 27 October 2026).
• EdAcc has a card that lists CC BY-SA (with some public domain material) and reports 40 hours of conversation (EdAcc card, read 27 October 2026).
• FLEURS covers 102 languages with about 12 hours of speech per language (Conneau et al., 2022). Its English subset is labelled en_us, and we found no UK regional label on its card. It suits a clean baseline more than an accent test.

Licence note: CC BY-SA asks for credit and for share-alike terms on adapted material. An 8 kHz copy of a recording may count as an adaptation. If you share such copies, check the licence terms first.

How do you test your own agent?

The studies above cannot tell you how your agent does, because the effect changes with the speech engine and the carrier route. The short version is to run the same scripted calls in clean audio and in phone-line audio, in the accents your callers have, and compare word error rate on names, numbers and postcodes. The steps, the noise levels and the checklist of cases are in how to test a voice agent on accents and background noise. For how mis-heard spelled letters affect caller verification, see how to test a voice agent's identity checks, and for how to plan the calls, see the call tester guide.

Where RingTru fits

RingTru's simulated callers ring your phone agent with UK accents, noisy phone-quality lines and interruptions, and the calls are scored against written rules (see accents and noisy lines). We have set out a method for a UK accent benchmark on phone audio. Round one has not run, so this article quotes no results of ours.

Sources were read on 27 October 2026. Figures are the papers' own, for their models and data. They do not predict your agent's results.