Method: Isolating Telephone Bandwidth in Speech Translation (2026)

Supporting method page for Phone Call Translation Accuracy. Published 2026-08-18.

Conflict of interest

LiveLingo is our product and one of the four systems measured. It placed first on both corpora. Readers should weigh that accordingly, and the limitations above are stated in full for the same reason. The headline bandwidth finding is a null result that applies to every system equally, including ours, and it removes a marketing claim we could otherwise have made about phone audio quality.

Degradation chain

Each 16 kHz mono source recording was transformed as follows: resample to 8 kHz with anti-alias filtering, encode to G.711 mu-law 8-bit, decode, resample back to 16 kHz mono PCM-16. The output container is identical in format to the input, so no downstream system can branch on sample rate or codec. The only difference between arms is the information a telephone channel discards.

Verification: energy above 3.4 kHz falls from 14-22% of total spectral energy in the wideband arm to approximately 3% in the narrowband arm, while RMS amplitude is unchanged to three decimal places (for example 0.0883 vs 0.0878). The comparison therefore measures bandwidth, not loudness.

Systems and scoring

Four systems processed both arms: LiveLingo; Google Cloud Speech-to-Text v2 with Translation v3; Azure Speech Translation; and a Whisper-large plus GPT-4o-mini baseline. Each output was scored 0-5 for comprehension fidelity by three independent LLM judges (GPT-4o-2024-11-20, Gemini 2.5 Flash, Claude Sonnet 4.6). The composite is the arithmetic mean of the three. The rubric penalises wrong-word substitution, dropped entities or numbers, punctuation that changes sentence type, and language-script mixing; it does not penalise register or valid synonyms.

Confidence intervals are bootstrap 95% intervals over paired per-clip differences, 5000 resamples with a fixed seed so the published figures are reproducible. Zero cells failed to score across 2,240 scored cells.

Same-day control

An earlier version of this analysis compared the narrowband arm against archived wideband results from six weeks prior. That comparison showed Azure losing 0.133 points, with a confidence interval excluding zero. Re-running the wideband arm on the same day through the same pipeline reduced that to 0.042 with an interval spanning zero. The apparent effect was product and vendor drift over the intervening weeks, not bandwidth. All figures published here use the same-day control.

Removed measurement artifacts

Raw edit distance between arms overstates degradation. Four artifacts were identified by reading transcripts rather than trusting aggregates, and each is normalised away before the “real drift” figure is computed:

• Hallucinated stock phrases. Whisper emits “subscribe to the channel”, “Legendas pela comunidade Amara.org” and YouTube URLs onto trailing silence, nondeterministically and in either arm. Source clips carry 1.5s of padding, so this fires frequently.
• Chinese script variance. Output flips between Simplified and Traditional at random. Normalised with OpenCC t2s; after conversion the affected pairs are identical.
• Arabic orthography. Differences are predominantly hamza and alef variants (أنا vs انا) plus dropped diacritics. These are spelling variants, not recognition errors.
• Word-level metrics on unspaced scripts. Chinese, Japanese and Thai have no word boundaries, so whitespace tokenisation yields one token per sentence and a single character change reads as 100% error. Those languages use character error rate.

Before these corrections the aggregate Whisper drift measured 5.7%, with individual cells as extreme as 69%. After correction it is 2.0%. The uncorrected figure would have been wrong in a way that flattered the article's original hypothesis.

Per-pair results, extended set

Comprehension fidelity, wideband → narrowband, 20 paired utterances per pair.

Statistical power

Minimum detectable effect at approximately 80% power, computed from the paired per-clip standard deviation at n=120: LiveLingo 0.058 points, Azure 0.015, Google 0.047, and the Whisper baseline 0.142. The design could therefore not have detected a true degradation smaller than those values. For three of the four systems that floor is well under the 0.04 threshold discussed, so the null is meaningful. For the Whisper baseline the per-clip variance is high enough that a real degradation of up to 0.14 points would not have been distinguished from noise, and its null should be treated as inconclusive.

Limitations

The source speech is synthesised, not human. Every clip in both corpora is text-to-speech, confirmed by a frame-level noise floor of essentially zero where a real microphone recording always carries one. Synthetic speech is hyper-articulated: it has canonical fricatives, no coarticulation or vowel reduction, no accent variation and no background noise. Because the mechanism band-limiting is predicted to damage is precisely fricative discrimination, synthetic speech is the most favourable possible case for the null result reported here. The finding should be read as an upper bound on robustness, and replication on human speech is the single most valuable next step.

The judging metric is coarse. 66% of wideband cells scored a perfect 5.0. Comprehension fidelity asks whether meaning survived, not whether quality declined, so a translation that degrades while remaining fully comprehensible scores identically in both arms by design. Source-transcript drift was measured separately for exactly this reason.

One measurement per arm. LLM judge panels are not fully deterministic, so a single paired pass cannot separate a small true effect from run-to-run variance. Repeating the comparison three or more times would tighten every interval reported here.

Codec only. The degradation chain models G.711 band-limiting and quantisation. It does not model packet loss, jitter, automatic gain control, noise suppression, comfort noise, or carrier-side transcoding, all of which occur on real calls. The correct reading is narrow: telephone bandwidth alone does not measurably reduce translation quality on clean speech. It is not a claim that phone calls are as easy as quiet in-person conversation, and the published CallHome and Switchboard error rates indicate strongly that they are not.

References

• Radford, Kim, Xu, Brockman, McLeavey, Sutskever. Robust Speech Recognition via Large-Scale Weak Supervision. arXiv:2212.04356, 2022. Table 2.
• ITU-T Recommendation G.711, Pulse code modulation of voice frequencies.
• Bauer & Fingscheidt. WTIMIT: The TIMIT Speech Corpus Transmitted Over The 3G AMR Wideband Mobile Network. LREC 2010.
• Coleman, J. Acoustic structure of consonants. University of Oxford Phonetics Laboratory.