How Accurate Is Phone Call Translation? 4 Systems Tested (2026)

Published 2026-08-18 · By Ron Villomo, Founder of LiveLingo · LiveLingo is one of the systems measured; full method and raw data

1. How accurate is phone call translation? The measured answer

We scored four systems on identical audio passed through a simulated phone leg. Each translation was rated 0-5 for comprehension fidelity by three independent frontier LLM judges, and the score below is the mean of the three, shown also as a percentage of the maximum. The rubric penalises wrong-word substitution, dropped entities or numbers, and language-script mixing; it does not penalise style or valid synonyms.

The spread between language pairs matters more than the spread between the top systems. Even the strongest system loses three points moving from German to Japanese, and every system loses far more on the harder corridor pairs than it loses to the phone line.

2. Does Apple Live Translation work on phone calls?

Apple Live Translation handles phone calls on recent iPhones and Samsung Galaxy AI offers a comparable in-call feature. Both are good at what they do, and for an iPhone-to-iPhone call in a supported language they are the most convenient option available because there is nothing to install.

Their limit is coverage, not quality. Both require a recent handset from that specific manufacturer, both are restricted to the language sets those makers ship, and critically both put the capability on your device rather than on the line. That is fine for calling a friend with the same phone. It does not help when you need to reach a clinic switchboard, a landlord, a hotel desk, or a supplier whose office phone is a decade old. Those calls are the reason most people go looking for a translated-call product in the first place.

3. Why is phone call audio quality bad, and does it reduce accuracy?

A phone call is encoded with ITU-T G.711, which samples at 8 kHz and carries roughly 300 Hz to 3400 Hz. The discarded band is not empty. Acoustic phonetics work from the University of Oxford's phonetics laboratory puts the energy peaks of /s/ at roughly 4500 Hz and 7500 Hz, entirely above the telephone ceiling, which is also why band-limited /s/ is systematically misheard as /sh/. The physics predicts phone audio should be harder. ITU-T G.711 · University of Oxford Phonetics Laboratory

To test whether that prediction survives to the output, we produced two versions of every clip. One is the original 16 kHz recording. The other is the same file resampled to 8 kHz with anti-alias filtering, quantised with G.711 mu-law, then decoded back into the original 16 kHz container, so both files are identical in format and no system can branch on sample rate. Energy above 3.4 kHz falls from 14-22% of the signal to about 3%, while RMS loudness is unchanged, so the comparison measures bandwidth and not volume.

Every system landed within 0.04 points of itself. LiveLingo scored 97.2% on microphone audio and 96.9% over the phone line; Google 93.4% and 94.0%; Azure 91.7% and 91.8%; the Whisper baseline 88.1% and 88.9%. 88% of individual cells received an identical score in both arms. Telephone bandwidth destroys real information and modern translation absorbs nearly all of it.

4. Recognition drifts about 2%, and robustness varies 5x between systems

Quality holding steady does not mean nothing happened underneath. Comparing the source transcript each system produced from the wideband file against the one it produced from the phone-line file, Google Cloud Speech-to-Text v2 changed 0.4% of recognised words and the Whisper-large baseline changed 2.0%, with 96% and 81% of transcripts respectively coming back completely unchanged.

Raw comparison overstates this badly, and the reason is worth stating because it is how this kind of number goes wrong. Whisper hallucinates stock phrases onto trailing silence, emitting fixed subtitle-style boilerplate unpredictably in either arm. Chinese output flips between Simplified and Traditional script at random. Arabic differs mostly by hamza spelling. Before removing those, Whisper's apparent drift measured 5.7%; after, it is 2.0%.

The useful finding is that narrowband robustness differs by about 5x between recognisers, which is a far larger effect than the phone line itself. Arabic and Indonesian drifted most at around 5%, consistent with the fricative prediction, and even there the fidelity score did not move.

5. Does AI translation get worse on phone calls than in person?

Because the evidence usually cited for it measures something else. In the Whisper paper (Radford et al., 2022), wav2vec 2.0 Large and Whisper Large V2 both score an identical 2.7% word error rate on LibriSpeech clean read speech. On the CallHome telephone corpus they diverge to 34.8% and 17.6%. That gap is real, large, and routinely offered as proof that phone audio is the problem. Radford et al., 2022 · LDC CallHome

But CallHome is not merely narrowband. It is spontaneous, overlapping, unscripted conversation between people who know each other well, dense with contractions, false starts and crosstalk. LibriSpeech is one person reading a book aloud. Changing two variables at once cannot tell you which one mattered. Isolating the bandwidth finds almost nothing, which means the CallHome penalty is mostly the other variable.

That is the more useful conclusion, because those are the factors you can actually control. Taking turns cleanly, moving somewhere quieter, and letting each speaker finish will do more for a translated call than any concern about line quality.

6. What actually separates phone call translation apps

Since bandwidth costs nothing measurable and the top systems sit within three points of each other on easy pairs, the differences that matter are elsewhere: how far quality falls on your specific language pair, whether the product can reach the number you need to call, and how the call handles turn-taking.

Language pair is the biggest quality lever by a wide margin, so check your own corridor rather than a headline average. On reach, LiveLingo places the translated call from the app to any mobile or landline number, so the recipient answers an ordinary phone call and installs nothing. And on conduct, it is worth knowing whether a product announces itself honestly or clones your voice to hide that a translator is involved. livelingo.io/phone-call-translation

7. What this test covers, and what it does not

The numbers above are bounded by four specific design choices. Stating them is not a hedge on the results; it is the scope within which the results hold.

The practical reading: the ranking in section 1 is a clean like-for-like comparison, since every system received identical audio and identical judging. The bandwidth result in section 3 is the claim most sensitive to the corpus choice, and replication on spontaneous conversational speech is the next step. Full derivations, statistical power and the measurement artifacts we removed are on the full method and raw data.

Frequently asked questions

Full method, per-language results, statistical power, limitations and the measurement artifacts we removed are documented at livelingo.io/research/phone-line-bandwidth-method.