"Can it handle Spanish?" is one of the most common questions you'll get on a discovery call, and the honest answer is "yes, but not the way you're imagining." Multilingual voice agents are real and they close deals — a bilingual line in the right market is a genuine differentiator. But most operators quote it like it's a checkbox, then discover the hard parts on a live call. Here's what actually matters.
Three different problems hiding under one request
When a client says "handle other languages," they could mean any of three things, and they cost different amounts of work:
- Serve callers in a second language — the whole call happens in Spanish (or Portuguese, or Vietnamese). This is the common ask and the most valuable.
- Detect and switch language mid-call — the caller opens in English, then switches, and the agent follows. Harder, occasionally flaky, and often unnecessary — a language menu up front ("Para español, oprima dos") is more reliable than auto-detection.
- Handle accented speech in one language — the caller speaks English with a strong accent and the agent keeps mishearing. This is a transcription problem, not a language problem, and it's the one that quietly kills calls.
Scope which one the client actually needs before you quote. Most want #1.
The stack has three layers, and each has to speak the language
A voice agent is STT → LLM → TTS. All three have to support the target language, and they're the usual suspects when it breaks:
- STT (transcription) — Deepgram, and the models behind Vapi/Retell/Bland, support many languages but with uneven quality. Spanish and major European languages are solid; smaller languages and heavy code-switching are where accuracy drops. This layer is almost always the weak link.
- LLM — the modern models (GPT, Claude, Gemini class) are genuinely strong across major languages, so this is rarely your problem. Just make sure your prompt is written in — or explicitly instructs the agent to respond in — the target language.
- TTS — ElevenLabs and Cartesia both do excellent multilingual output; pick a native voice for the language, not an English voice speaking Spanish, which sounds subtly wrong and erodes trust. Audition it with a native speaker.
Accents: the problem nobody quotes for
Even in pure English, a strong accent — regional, or a non-native speaker — can wreck the transcription layer, and the LLM can only work with the garbled text it's handed. Symptoms: the agent asks callers to repeat, mishears names and numbers, or confidently books the wrong thing. Mitigations:
- Pick the STT model that handles your caller base. Test with real accented audio from the actual market, not your own clean voice.
- Lean on confirmation for high-stakes fields. Always read back names, phone numbers, dates, and dollar amounts and get a yes. This is good practice everywhere; with accents it's mandatory.
- Constrain where you can. Spelling capture for names, digit-by-digit confirmation for numbers, and a short escalation threshold (two failed attempts → human) beat any amount of prompt tweaking.
Test with real speakers, not Google Translate
The single biggest mistake is building the Spanish agent in English, running the prompt through a translator, and shipping it. Idioms, formality (tú vs. usted), and the cultural register are all wrong, and a native-speaking caller hears it instantly. Have a native speaker write or review the prompt and run the same edge-case scenarios — hard names, numbers, interruptions, accents — on a real phone in that language. Budget for this; it's the step that makes the difference.
Where the money is
Bilingual English/Spanish lines are a strong wedge in the US — home services, medical, legal intake, and property management all have large Spanish-speaking caller bases that competitors serve badly or not at all. You can often charge a premium for it because the alternative is the client staffing a bilingual receptionist, which is expensive and hard to hire. Just price the build honestly — a second language is a second prompt, a second voice, a second round of testing, not a toggle — and you'll deliver something that actually works on the phone instead of a demo that falls apart the moment a real caller opens in Spanish.