Which languages the voice agent follows, and how to test yours
How GPT-Live 1 decides what language to answer in, which languages OpenPhonex has tested, what "not tested" means, and how to test your own language before going live.
This page is about the GPT-Live 1 conversation engine. A Standard (speech-to-text, model, text-to-speech) agent is configured for one language and qualified per provider; see How OpenPhonex handles provider languages and voices.
How language following works
GPT-Live 1 has no language setting. No API parameter tells it which language to speak, and none restricts it to one. The model listens to what the caller says and replies in the language it believes it heard.
- Voice and language are independent. The voice you pick sets the timbre of the reply, not the language. Any of the fourteen GPT-Live voices can answer in any language the model recognises. In our tests the same voice answered German callers in German and Tamil callers in Tamil.
- Language following is automatic and not enforceable. You cannot pin an agent to one language. An instruction such as "only ever speak English" is a preference the model may or may not follow, not a rule it obeys. In our tests it did not hold.
There is nothing to configure, and so nothing to misconfigure. Plan around this instead: the agent replies in whatever language it believes the caller used, every time.
What OpenPhonex has tested
We ran GPT-Live 1 through spoken test sessions in several languages, using synthetic callers speaking real, synthesised sentences. A session passed when the reply stayed in the caller's language, addressed the caller's actual question, and was coherent and on topic.
| Language | Result | Sessions |
|---|---|---|
| English | Reliable: in-language and on topic every time | 4 |
| Estonian | Reliable: in-language and on topic every time | 4 |
| Russian | Reliable: in-language and on topic every time | 4 |
| German | Reliable: in-language and on topic every time | 4 |
| Hindi | Reliable: in-language and on topic every time | 4 |
| Tamil | Reliable: in-language and on topic every time | 4 |
| Sinhala | Unreliable: 9 of 60 sessions usable | 60 |
Tested 13 September 2026 against gpt-live-1. Results apply to that model
version; we re-run the same sessions when OpenAI ships a new one.
All fourteen GPT-Live voices remain available for every language. OpenPhonex does not restrict voice choice by language, because voice is a timbre choice and has no bearing on how GPT-Live 1 follows language. In the Voice Library they appear under every language filter in a Multilingual group.
What "not tested" means
A language missing from the table is not tested, not unsupported. We have not run it through this process, so we have no result to report either way. If your business takes calls in a language that is not listed, test it yourself before relying on it.
What a failure looks like
Sinhala is the one language where we found a real, repeatable problem, and it is a useful example of what "unreliable" means in practice.
Speech recognition frequently misheard the caller's Sinhala sentence, and the reply did not just get the content wrong: it sometimes came back in a completely different language that nobody spoke. Across 60 sessions, replies landed in Tamil, Hindi, Marathi, Tagalog, Japanese, Portuguese and Hebrew at least once each.
One example: a caller asked, in Sinhala, about their loan payment due date. The recogniser heard Japanese, and the model replied fluently in Japanese, apologising and asking the caller to repeat the question in Japanese, a language the caller never spoke.
This is not tied to one voice. Identical audio produced different results on different sessions, so it is session-to-session randomness rather than something a voice or setting change would fix. German, Tamil and Hindi produced no such failures, so this looks specific to Sinhala rather than a general non-English or non-Latin-script problem.
How to test your own language
Before routing real traffic to a language, test it in the Simulations Lab. It runs your actual agent configuration against a synthetic caller using real speech, so you see the same language-following behaviour a live call would produce.
Current limit
The synthetic caller can speak Sinhala and Tamil for any agent today. For other languages it needs a caller voice on the agent's own speech provider, which GPT-Live 1 agents do not have, so the caller stays silent. A caller voice for every language is planned; until then, test other languages with a real test call to your number.
Open the Simulations Lab for the agent you want to test.
Create a scenario and set its language field to the language you are testing.
Author a persona for the synthetic caller: an identity, a goal, and how it behaves. The persona's lines are synthesised as real speech in the scenario's language, not typed text.
Run the scenario. The Lab drives your agent through the call as a live caller would, using a real synthetic voice for that language.
Read the transcript. Check whether replies stayed in the caller's language and addressed what was actually asked.
Repeat with a few sentences. A single utterance will not catch session-to-session randomness; our own Sinhala testing needed many runs to see it.
The Simulations Lab does not yet flag a language mismatch automatically. Check the transcript yourself; it will not surface as an error elsewhere in the run.
If it fails
If your tests come back unreliable for a language, there is no setting that fixes GPT-Live 1's language following. The alternative is the Standard engine: a dedicated speech-to-text provider and a separate text-to-speech provider, chosen and qualified per language, instead of one model handling both recognition and reply.
OpenPhonex qualifies Standard languages per provider, and coverage varies: some languages have a qualified speech-to-text provider today, others do not yet. If yours is not qualified on the Standard engine, talk to us before committing it to production.