One agent, every language: how multilingual calling works

A caller dials in and speaks Vietnamese. The agent greets them in Vietnamese, retrieves the right answer from your knowledge base, and responds — without any routing rule, language configuration, or extra setup on your part. The same agent, the same phone number, handles an English caller thirty seconds later.

Multilingual support in Auralius is not a feature you turn on. It is how the pipeline works by default.

How language detection works

The speech-to-text engine runs continuous language detection alongside transcription. As the first words arrive from the caller, the model identifies the language from phonetic and lexical patterns — typically within the first one to two seconds of speech. The detected language is attached to the transcript before it reaches the language model and the text-to-speech engine.

This means the agent does not wait for the caller to finish a sentence before knowing what language they are speaking. Detection happens in the same stream as transcription, adding no perceptible delay.

The agent responds in the same language

Once the language is detected, the full pipeline — retrieval, reasoning, and speech synthesis — operates in that language. The language model generates a response in the detected language. The text-to-speech engine, which uses a multilingual synthesis model, speaks it back with natural cadence and pronunciation.

If the caller switches languages mid-call — common in multilingual households or markets where callers code-switch — the detection updates turn by turn. The agent adapts without the caller noticing any transition.

Your knowledge base works across languages

The retrieval-augmented generation layer handles cross-language retrieval. If your knowledge base is written in English and a caller asks in Vietnamese, the retrieval finds the relevant passages and the language model synthesizes the answer in Vietnamese. You do not need to maintain separate knowledge bases per language.

For the most accurate results in a specific language, including that language in your knowledge base content is always better. But the system does not require it to function.

When to create separate agents per language

Auto-detection works well for businesses with mixed caller populations. For use cases where a single language dominates — a Japanese-market clinic, a Vietnamese e-commerce support line — creating a dedicated agent with a voice selected for that language gives the most natural result. You can assign a voice that is native to your primary language rather than relying on the multilingual model's general-purpose output.

The choice comes down to your caller mix. If it is homogeneous, use a dedicated agent. If it is diverse, auto-detection handles it without any extra effort.

Market coverage

Auralius is built for North American and Asian markets first. The speech recognition and synthesis pipeline covers the major languages of those markets — English, Vietnamese, Japanese, Mandarin, Korean, and others — with more being added as the platform scales. European languages are supported at the platform level; EU-specific compliance features are available for tenants operating in that region.