Each agent has a primary language (language). It sets the speech recognition language and tells the voice which language to speak. An agent can also accept several languages on one call with languages, which switches transcription to multi-language mode.
Set the language#
Click the language picker in the toolbar, for example English (AU), and pick a locale. This sets language on the agent. The default is en-AU.
The picker offers 73 codes:
| Region | Codes |
|---|---|
| English | en-AU, en-US, en-GB, en-NZ, en-IN, en-CA, en-IE, en-ZA, en-SG, en-PH |
| Spanish, French, German | es-ES, es-419, es-US, fr-FR, fr-CA, fr-BE, de-DE, de-AT, de-CH |
| Other Western European | it-IT, pt-BR, pt-PT, nl-NL, nl-BE, ca-ES |
| East Asian | ja-JP, ko-KR, zh-CN, zh-TW, zh-HK |
| South Asian | hi-IN, bn-IN, ta-IN, te-IN, mr-IN, gu-IN, kn-IN, ml-IN, pa-IN, ur-PK |
| Middle Eastern | ar-SA, ar-AE, ar-EG, he-IL, fa-IR, tr-TR |
| Eastern European | ru-RU, uk-UA, pl-PL, cs-CZ, sk-SK, hu-HU, ro-RO, bg-BG, hr-HR, sr-RS, sl-SI, el-GR |
| Nordic and Baltic | sv-SE, da-DK, no-NO, fi-FI, lt-LT, lv-LV, et-EE |
| South-East Asian and African | id-ID, ms-MY, th-TH, vi-VN, fil-PH, sw-KE, af-ZA |
| Any | multi |
The full code, region included, is sent to the speech recognizer. Pick the caller's region: en-AU for Australian callers.
Speak the right language#
Choosing a language does not translate your prompt. Write the prompt and the welcome message in the language the agent should speak, or tell the model which language to reply in.
The voice receives the two-letter language code (for example en from en-AU) for both voice tiers. Pick a voice that speaks the language well and test it with Test Audio. Preview with Test Audio.
Multilingual calls#
When callers may use more than one language, select them under Realtime Transcription Settings, Languages callers may speak. The field starts with the primary language. Search and tick each extra language.
| Setting | Result |
|---|---|
| One language (the default) | Transcription runs in that language. |
| Two or more different languages | Transcription runs in multi-language mode and follows the caller as they switch. |
language: "multi" |
Same as two or more languages. |
The selection is stored in languages. Selecting only the primary language stores an empty list.
Add a line to the prompt so replies follow the caller: "Reply in the language the caller uses. If they switch, switch with them."
By transcription provider#
| Provider | Multi-language behaviour |
|---|---|
| Deepgram (default) | nova-3 in multi-language mode. With Vocabulary set to Medical, multi-language mode uses nova-3 rather than nova-3-medical. |
| AssemblyAI | The multilingual streaming model, also used for any single non-English language. |
| Soniox | Receives each language as a hint. |
| Azure | Receives the list of languages. |
| OpenAI | Detects the language. |
Providers other than Deepgram are chosen with Transcription Mode, Custom. See Speech recognition.
Fields#
| Field | Type | Default | Description |
|---|---|---|---|
language |
string | en-AU |
Primary locale. One of the codes above. |
languages |
string array | not set | Languages callers may speak. More than one distinct language turns on multi-language transcription. |
curl -X PATCH https://api.tried.com.au/update-agent/agent_3f1c9a7b2d4e6f8091a2b3c4 \
-H "Authorization: Bearer $TRIED_API_KEY" \
-H "Content-Type: application/json" \
-d '{ "language": "en-AU", "languages": ["en-AU", "zh-CN", "vi-VN"] }'Tips#
- Boosted keywords help with names in any language. See Speech recognition.
- The runtime uses a multilingual turn detector by default, so end-of-turn timing applies to non-English calls too.
- Post-call analysis runs on the transcript in whatever language was spoken. Write analysis field descriptions that say which language the output should be in.