Research · Published · By Ron Villomo
Can GPT-Live-1 work as a simultaneous interpreter?
We prompted OpenAI's full-duplex voice model to interpret 7.5 minutes of continuous speech and measured how far behind the speaker it ran. It holds the role perfectly. It just stops keeping up.
Disclosure. We build LiveLingo, one of the systems measured here. Every number comes from the same audio, the same clock and the same alignment code, and the per-clip data is published so the comparison can be checked rather than trusted.
Key findings · 452 s per run, 90 sentences, 4 systems, 2 language pairs
- GPT-Live-1 never breaks character. Across 27 minutes of interpreted audio it never once answered the speaker instead of translating them.
- It falls progressively behind. In English→Spanish its completion lag climbs from 5.7 s to a 10–14 s plateau within three minutes and stays there.
- In English→Japanese it collapses. Only 17% of sentences were translated within 30 s, against 98% for the same model into Spanish. The median output landed 50.7 s late and 20 of 90 sentences were never translated. Reproduced twice.
- This is not an OpenAI problem. OpenAI's own dedicated model, gpt-realtime-translate, held 3.33 s flat with 98% coverage on the identical Japanese audio, same account, same day.
Headline results
Completion lag is the moment a sentence's translation is fully delivered, minus the moment the speaker finished saying it. Every system received identical audio at identical real-time pace.
| System | en→es lag | en→ja lag | en→es coverage | en→ja coverage | Across session |
|---|---|---|---|---|---|
| LiveLingo | 0.94 s | 0.88 s | 100% | 100% | flat |
| Gemini 3.5 Live Translate | 1.51 s | 1.92 s | 100% | 99% | flat |
| gpt-realtime-translate | 2.65 s | 3.33 s | 99% | 98% | flat |
| GPT-Live-1 (prompted) | 11.72 s | n/a | 98% | 17% | rises to 10–14 s |
GPT-Live-1's Japanese lag is withheld deliberately: when 20 of 90 sentences never arrive, a median describes only the part that survived. Coverage is the honest measure there.
How far behind does GPT-Live-1 run?
The three purpose-built translation paths are flat: lag at minute seven is within a few hundred milliseconds of lag at minute zero. GPT-Live-1 starts at 5.7 s and settles between 10 and 14 s. The shape matters, because it is not unbounded. Lag rises for about three minutes, then plateaus. A listener joining at minute six hears a translation roughly twelve seconds behind the room.
| System | m0 | m1 | m2 | m3 | m4 | m5 | m6 | m7 |
|---|---|---|---|---|---|---|---|---|
| LiveLingo | 0.8 | 1.0 | 0.8 | 0.8 | 0.9 | 1.0 | 0.9 | 0.8 |
| Gemini 3.5 Live Translate | 1.5 | 1.6 | 1.6 | 1.6 | 1.5 | 1.5 | 1.4 | 1.4 |
| gpt-realtime-translate | 3.3 | 2.8 | 2.1 | 2.8 | 2.4 | 2.0 | 2.8 | 2.4 |
| GPT-Live-1 (prompted) | 5.7 | 10.4 | 14.1 | 10.7 | 13.4 | 10.0 | 11.4 | 13.7 |
What happens in English to Japanese?
Japanese is where the behaviour stops being a latency question and becomes a delivery question. Rather than quote a lag on sentences that may never have arrived, we counted coverage: of 90 spoken sentences, how many received a translation within 30 seconds of being said.
- Gemini 3.5 Live Translate: 99%
- gpt-realtime-translate: 98%
- GPT-Live-1 into Spanish: 98%
- GPT-Live-1 into Japanese: 17%
The median GPT-Live-1 output arrived 50.7 s after the sentence it translated, and 82% of its output landed more than 25 s late. It did not drain a backlog once the audio stopped: it ended roughly when the input did and abandoned what it had not reached. We ran it twice, getting 17% coverage both times with median offsets of 54.6 s and 50.7 s.
Is this OpenAI, or is it prompting a dialogue model?
This is the question that decides whether the finding means anything, so we ran the control. OpenAI ships a dedicated streaming translation model, gpt-realtime-translate, released 7 May 2026 at $0.034 per minute. We put it on exactly the same Japanese stream, from the same account, on the same day.
It held 3.33 s flat, covered 98% of sentences within 30 s, and had zero output landing more than 25 s late. OpenAI's stack handles Japanese fine. The collapse belongs to the prompted full-duplex dialogue model specifically, not to the vendor.
That points at a mechanism rather than a bug. A conversational model is built to hold a turn at a natural speaking rate, and it neither compresses nor drops when it falls behind. Interpretation has the opposite requirement: the speaker keeps going regardless, so a system that cannot keep pace must summarise, skip, or accept unbounded delay. GPT-Live-1 chose delay, then silence.
When should you use GPT-Live-1 anyway?
For what it was built for. On Artificial Analysis' Speech-to-Speech index it scores 81.5, second overall, with 1.34 s to first audio. It is a strong full-duplex voice agent, it interrupts and backchannels naturally, and it held the interpreter role without a single lapse across 27 minutes, which is genuinely hard, and which the dedicated translation models do not even have to attempt.
For a short exchange, say a few sentences or a question at a counter where both parties pause, the drift never has time to accumulate. Our Spanish session was already 10 s behind by minute three, but minute zero measured 5.7 s and the first exchange felt immediate.
For continuous speech, a meeting or a lecture or a call where one side talks for minutes at a time, use a path built for it, whether that is OpenAI's own gpt-realtime-translate, Google's Gemini 3.5 Live Translate, or ours.
Methodology
The source is the 30-sentence corpus published with our 2026 comprehension benchmark: synthesised English speech across five conversational domains, concatenated with 0.5 s gaps and repeated three times to make a 452-second continuous stream. Reusing that corpus means the reference translations already exist, so every system is scored against the same anchors.
- Source timing. Word-level timestamps over the source audio, so each sentence's true end time is known independently of any system under test.
- Alignment. Each output sentence is matched to the source sentence whose published reference translation it overlaps most, by monotonic dynamic programming with a causality constraint. Japanese matches on kanji and katakana bigrams.
- Pacing. Audio streamed at real-time pace against absolute deadlines; the sender finished within 39 ms of the source duration.
- Metric. Completion lag = last committed token of a sentence's translation, minus that sentence's end in the source.
Scope
| Language pairs | English→Spanish and English→Japanese, plus a Chinese→English spot check |
|---|---|
| Source audio | Synthesised speech, single speaker, no overlap or background noise |
| Session length | 452 s continuous, plus a 151 s run |
| Runs | One per system per pair; the Japanese collapse reproduced twice |
| Measured | Completion lag and 30-second coverage |
| Not measured | Translation fidelity, voice quality, prosody, cost per session |
| Output form | LiveLingo commits text; the other three emit speech. Timing is compared at the moment the translation becomes available. |
Download the data
Every number on this page is derived from these files, released under CC-BY 4.0.
- results.json (headline metrics, per-minute curves and full methodology for all 8 runs)
- per-clip.json (720 rows: every sentence, its source timing, the translation produced and its completion lag)
Questions
Can GPT-Live-1 translate in real time?
Yes, and it follows the instruction reliably. Prompted to act as a simultaneous interpreter it translated continuously for 7.5 minutes without once answering the speaker. The limitation is pace, not capability: completion lag rose from 5.7 s to a 10–14 s plateau in English→Spanish.
How much delay does GPT-Live-1 add when interpreting?
GPT-Live-1 took a median of 11.72 s to finish delivering a sentence in English→Spanish, rising through the session from 5.7 s at minute zero. On identical audio, gpt-realtime-translate measured 2.65 s, Gemini 3.5 Live Translate 1.51 s and LiveLingo 0.94 s, all flat across the same 452 seconds.
Why does GPT-Live-1 fail on Japanese?
It falls far enough behind that it stops catching up. Only 17% of sentences were translated within 30 s, against 98% for the same model into Spanish, and it stopped emitting about 7 s after the audio ended leaving 20 of 90 sentences untranslated. OpenAI's dedicated gpt-realtime-translate covered 98% at 3.33 s flat on the same audio.
Should I use GPT-Live-1 or gpt-realtime-translate for live translation?
Use gpt-realtime-translate for translation. It is purpose-built, costs $0.034 per minute against $0.05 per minute for the GPT-Live-1 voice layer plus backend model usage, and it held a constant lag in both pairs tested. Use GPT-Live-1 when you want a conversational voice agent.
Which real-time translation API has the lowest latency?
On this corpus, LiveLingo at 0.94 s (en→es) and 0.88 s (en→ja), then Gemini 3.5 Live Translate at 1.51 s and 1.92 s, then gpt-realtime-translate at 2.65 s and 3.33 s. All three hold that figure across a full session; GPT-Live-1 prompted as an interpreter does not.
Does a full-duplex voice model work as an interpreter?
It holds the role but not the pace. A conversational model speaks at a natural rate and neither compresses nor drops when it falls behind, while an interpreter must keep up with a speaker who does not wait. GPT-Live-1 never broke character across 27 minutes, yet accumulated delay until it plateaued at 10–14 s in Spanish and left 20 of 90 Japanese sentences untranslated.
Cite this benchmark
LiveLingo Research (2026). GPT-Live-1 as a Simultaneous Interpreter: Latency and Coverage (2026). https://www.livelingo.io/research/gpt-live-1-interpreter-test
@misc{livelingo2026gptlive1interpretertest,
author = {{LiveLingo Research}},
title = {GPT-Live-1 as a Simultaneous Interpreter: Latency and Coverage (2026)},
year = {2026},
howpublished = {\url{https://www.livelingo.io/research/gpt-live-1-interpreter-test}},
note = {Dataset (CC-BY 4.0): https://www.livelingo.io/research/gpt-live-1-interpreter-test/results.json},
}