This page is the canonical reference for which languages PolyAI supports and which models cover them across LLM, speech recognition (ASR), and text-to-speech (TTS).
For setup and runtime behavior, see Multi-language. For model selection, see Model and Raven.
At a glance
| Layer | Coverage | Notes |
|---|
| LLM response languages | 73 (table below) | Compiled list of BCP 47 codes accepted by the platform. |
| Raven (recommended LLM) | 24 languages | Conversation-tuned. Some languages outside this list (for example Danish) require a third-party LLM. |
| Third-party LLMs | Provider-dependent | Pick a model that supports your target language; channel availability varies. |
| ASR (speech-to-text) | Provider-dependent | Routed automatically per language with fallback. |
| TTS (text-to-speech) | Provider-dependent | Voice availability varies by provider and language. |
LLM language coverage
Raven (recommended)
Raven is purpose-built for customer service across voice and chat. Raven 3.5 supports the following 24 languages:
Arabic, Bulgarian, Cantonese, Croatian, Czech, Dutch, English, French, German, Greek, Hindi, Hindi (Romanized/Hinglish), Italian, Japanese, Korean, Mandarin (PRC), Mandarin (Taiwan), Polish, Portuguese (Brazil), Portuguese (Portugal), Serbian, Spanish (US), Swedish, Turkish.
Strongest performance relative to general-purpose models: Cantonese, Italian, Korean, Mandarin (China), Mandarin (Taiwan), Spanish (US).
You can keep prompts and knowledge in English and set the response language to your target language – Raven responds consistently in the target language. Quality improves further if you translate prompts and add examples in the target language.
Third-party LLMs
When you select an OpenAI or Amazon Bedrock model on the Model page, language coverage matches that provider’s official support. Use a third-party LLM when:
- Your target language is not on the Raven list above, or
- You need a capability only the third-party model provides.
Full list of supported response languages
PolyAI accepts the BCP 47 codes below as response languages. Pass the code in the UI or via conv.set_language().
| Language | Code |
|---|
| Albanian | sq-AL |
| Amharic | am-ET |
| Arabic | ar |
| Armenian | hy-AM |
| Bengali | bn-BD |
| Bosnian | bs-BA |
| Bulgarian | bg-BG |
| Burmese | my-MM |
| Cantonese | yue-Hant-HK |
| Catalan | ca-ES |
| Chinese (China) | zh-CN |
| Chinese (Taiwan) | zh-TW |
| Croatian | hr-HR |
| Czech | cs-CZ |
| Danish | da-DK |
| Dutch (Belgium) | nl-BE |
| Dutch (Netherlands) | nl-NL |
| English (Australia) | en-AU |
| English (Canada) | en-CA |
| English (New Zealand) | en-NZ |
| English (Singapore) | en-SG |
| English (UK) | en-GB |
| English (US) | en-US |
| Estonian | et-EE |
| Finnish | fi-FI |
| French (Belgium) | fr-BE |
| French (Canada) | fr-CA |
| French (France) | fr-FR |
| Georgian | ka-GE |
| German (Germany) | de-DE |
| Greek | el-GR |
| Gujarati | gu-IN |
| Hebrew | he-IL |
| Hindi | hi |
| Hungarian | hu-HU |
| Icelandic | is-IS |
| Indonesian | id-ID |
| Italian (Italy) | it-IT |
| Japanese | ja-JP |
| Kannada | kn-IN |
| Kazakh | kk-KZ |
| Korean | ko-KR |
| Latvian | lv-LV |
| Lithuanian | lt-LT |
| Macedonian | mk-MK |
| Malay | ms-MY |
| Malayalam | ml-IN |
| Marathi | mr-IN |
| Mongolian | mn-MN |
| Norwegian | nb-NO |
| Persian (Farsi) | fa-IR |
| Polish | pl-PL |
| Portuguese (Brazil) | pt-BR |
| Portuguese (Portugal) | pt-PT |
| Punjabi | pa-IN |
| Romanian | ro-RO |
| Russian | ru-RU |
| Serbian | sr-RS |
| Slovak | sk-SK |
| Slovenian | sl-SI |
| Somali | so-SO |
| Spanish (Spain) | es-ES |
| Spanish (US) | es-US |
| Swahili | sw-KE |
| Swedish | sv-SE |
| Tagalog (Filipino) | tl-PH |
| Tamil | ta-IN |
| Telugu | te-IN |
| Thai | th-TH |
| Turkish | tr-TR |
| Ukrainian | uk-UA |
| Urdu | ur-PK |
| Vietnamese | vi-VN |
Serbian uses sr-RS (Republic of Serbia). If you were previously using the non-standard sr-SP code, update your project configuration to sr-RS.
Models
For full model descriptions and selection guidance, see Model.
LLM models
| Provider | Model | Channel | Regions | Notes |
|---|
| PolyAI | Raven 3.5 | Voice + Chat | All | Recommended. Auto-reasoning, out-of-domain detection, custom style following, built-in safety. |
| PolyAI | Raven V3 | Voice | All | Legacy. Superseded by Raven 3.5. No chat support. |
| PolyAI | Raven V2 | Voice | All | Legacy. Retained for existing deployments. |
| OpenAI (Azure / OpenAI direct) | GPT-5 | Voice + Chat | All | Strong reasoning. |
| OpenAI (Azure / OpenAI direct) | GPT-5 mini | Voice + Chat | All | Lower latency for mid-complexity workloads. |
| OpenAI (Azure / OpenAI direct) | GPT-5 nano | Voice + Chat | All | Fast, lightweight responses. |
| OpenAI (Azure / OpenAI direct) | GPT-5 chat | Chat | All | Optimized for extended dialogue. |
| OpenAI (Azure) | GPT-5.2 chat | Chat | All | Latest chat-optimized model. |
| OpenAI (Azure / OpenAI direct) | GPT-4.1 | Voice + Chat | All | Strong reasoning with improved cross-task performance. |
| OpenAI (Azure / OpenAI direct) | GPT-4.1 mini | Voice + Chat | All | Cost-effective, latency-focused. |
| OpenAI (Azure / OpenAI direct) | GPT-4.1 nano | Voice + Chat | All | Minimal compute, high throughput. |
| OpenAI (Azure / OpenAI direct) | GPT-4o | Voice + Chat | All | Balanced reasoning, speed, and cost. |
| OpenAI (Azure / OpenAI direct) | GPT-4o mini | Voice + Chat | All | High-volume / everyday workloads. |
| OpenAI | GPT realtime / realtime-mini | End-to-end voice | All | Speech-to-speech; behaves differently from text LLMs. |
| Amazon Bedrock | Claude Sonnet 4 | Voice + Chat | US-1, EU-W-1, UK-1 | Strong reasoning and safety alignment. |
| Amazon Bedrock | Claude 3.5 Haiku | Voice + Chat | US-1 only | Predictable tasks with strong safety alignment. Not currently configured in EU-W-1 / UK-1. |
| Amazon Bedrock | Nova Micro | Voice + Chat | All | Efficient general-purpose performance. |
| Custom | Fine-tuned OpenAI | Voice + Chat | All | Per-account fine-tuned OpenAI deployments. |
| Custom | Bring your own model | Voice + Chat | All | See Bring your own model. |
ASR providers
ASR providers wired into the platform today:
- Deepgram
- Google Cloud Speech-to-Text (v1 and v2)
- NVIDIA Riva
- Amazon Transcribe
- OpenAI (Whisper)
- Fano
- NVIDIA NeMo
The platform routes requests to the best-fit provider per language and use case, with automatic fallback. See Speech recognition.
TTS providers
TTS voice availability varies by provider and language. Browse what’s available per language in the Voice library.
- ElevenLabs
- Amazon Polly
- Azure Speech
- Cartesia
- Google Cloud Text-to-Speech
- Hume
- MiniMax
- Neuphonic
- OpenAI
- PlayHT
- Rime
- Custom TTS integrations
Choosing a language and model
- Pick the response language from the table above using its BCP 47 code.
- Check Raven coverage – if your language is on the 24-language Raven list, Raven 3.5 is the recommended LLM.
- If Raven does not cover it (for example Danish), select a third-party LLM in Voice configuration or Chat configuration. Verify the chosen model officially supports the language.
- Confirm regional availability for Bedrock models if you are deploying outside US-1.
- Confirm a voice exists for the language in the Voice library. Prefer native voices over multilingual fallbacks.
- Configure ASR – defaults usually work; for domain terms see Speech recognition and ASR biasing.
Related pages
