--- title: "Multilingual voice agents" description: "One Speko voice agent speaks up to four languages and switches when the caller does. Each language uses its own speech-to-text, LLM and text-to-speech, chosen from an independent benchmark." canonical: "https://speko.ai/multilingual-voice-agents" last-updated: "2026-11-11" --- > ## Speko page index > The complete index of every page on this site is at: https://speko.ai/llms.txt > Read it before exploring further. It lists the exact Markdown URL for every canonical HTML page. # Multilingual voice agents Speko Agents One agent speaks up to four languages and switches when the caller does. Speko is a voice agent platform for calls in more than one language. One agent holds up to four languages. When the caller switches, Speko swaps the speech-to-text, LLM, text-to-speech and prompt to match. [Get started free](https://platform.speko.ai/sign-in) [Book a demo](https://speko.ai/book) Agent stack Caller speaks English **Speech-to-text** assemblyai:universal-3-5-proWord error rate 2.0% **LLM and prompt** Prompt and replies in English **Text-to-speech** elevenlabs:eleven_v4Arena Elo 1708 The top-ranked models Speko can route to in each language, from its public benchmark. The default stack of an agent comes from the same data, and you can pin your own. ## One agent, several languages [Multilingual agents](https://docs.speko.ai/guides/multilingual-agents) - An agent holds up to four languages, each with a different primary language tag. When the caller switches language, all the settings below switch too. Set per language - Speech-to-text - LLM - Text-to-speech and voice - System prompt - Greeting, inbound and outbound - Voicemail line - Workflow text ## The right model per language [Speko benchmarks](https://speko.ai/benchmarks) - Speko's benchmark measures speech-to-text and text-to-speech models per language. Speko computes the default stack of an agent from the same data. Each language can also pin its own providers. Model EN (English) ES (Spanish) DE (German) FR (French) AR (Arabic) FIL (Filipino) NB (Norwegian) HI (Hindi) TA (Tamil) TE (Telugu) KO (Korean) ZH (Chinese (Mandarin)) JA (Japanese) TH (Thai) LT (Lithuanian) ID (Indonesian) SW (Swahili) VI (Vietnamese) deepgram:nova-3 soniox:stt-rt-v5 OpenAI:GPT-4o Transcribe bench openai:gpt-4o-mini-transcribe cartesia:ink-whisper Alibaba:Qwen3-ASR bench Google:Chirp 3 bench openai:gpt-transcribe xai:grok-stt assemblyai:universal-3-5-pro ElevenLabs:Scribe v2 bench Gladia:Solaria-1 bench ElevenLabs:Scribe v1 bench Smallest AI:Pulse Pro bench Alibaba:Qwen-ASR bench azure:MAI-Transcribe-2 Deepgram:Nova-2 bench modulate:velma-2-stt-streaming AssemblyAI:Universal-2 bench Azure:Fast Transcription bench fish:transcribe-1 fish:transcribe-1-pro gemini:gemini-3.5-transcribe gradium:default OpenAI:Whisper-1 bench meta:muse-voice-transcribe-1.0 alibaba:qwen3-asr-flash-realtime elevenlabs:scribe_v2_realtime AssemblyAI:Universal (batch tier) bench cartesia:ink-2 Groq:Whisper Large v3 bench hamsa:s3 hopper:nemotron-asr modulate:velma-2-stt-streaming-english-v2 munsit:munsit munsit:munsit-en-ar nari:qwen3-asr nari:qwen3-asr-fast omi:omi-medical-1 smallest:pulse Hover a cell for its number. worse better 15 of 58 are only measured in English Their rank in any other language is unknown — including the model that sits at the top of the English table. 6 different models win across 18 languages No single model is best everywhere, so the right pick changes with the language your users speak. ## Test it before callers do [Test before you deploy](https://docs.speko.ai/agents/testing) - The Switches language check passes when the agent replies mainly in the target language. You can also pin one language for a call, with no mid-call switching. ## Limits - Text a person wrote, such as a greeting or a workflow line, is spoken as written; write it in each language. - On a speech-to-speech (Realtime) call nothing switches mid-call. - Up to four languages per agent. ## Notes - Every fact is from the Speko docs page linked beside its section, checked November 10, 2026. ## Questions **Can one voice agent handle English and Spanish callers?** Yes. One Speko agent holds up to four languages. When the caller switches language, Speko swaps the speech-to-text, text-to-speech, LLM and system prompt to match. The agent then speaks the greeting and workflow text for that language. **How does Speko pick speech models for each language?** Each language can pin its own providers and voice. By default, Speko computes the stack from its benchmark, which measures speech models language by language. The benchmark is public at speko.ai/benchmarks. [Get started free](https://platform.speko.ai/sign-in) [Book a demo](https://speko.ai/book) You get $10 in credit when you sign up. Calls then cost $0.09 per minute. Related - [Voice agent testing with simulated callers](https://speko.ai/voice-agent-testing) - [Voice agent monitoring for every real call](https://speko.ai/voice-agent-monitoring) - [Voice agent workflow builder](https://speko.ai/voice-agent-workflow-builder) - [Voice agent API and SDKs](https://speko.ai/voice-agent-api) - [Voice agent platforms compared](https://speko.ai/voice-agent-platforms) - [Vapi alternatives](https://speko.ai/vapi-alternatives) - [Voice agents by business](https://speko.ai/solutions) - [Speko pricing](https://speko.ai/pricing)