Multilingual TTS for Voice Agents: Hindi, Spanish, Arabic

ElevenLabs Multilingual v2 and Flash v2.5, Cartesia Sonic 3.6, Google Chirp 3 HD and Murf all list Hindi, Spanish and Arabic voices. Azure Speech lists 16 Arabic locales, Sarvam covers Hindi and English-Hindi, and Deepgram Aura-2 covers Spanish. Dograh, the open-source AI voice agent platform, runs each of them, self-hosted or on Dograh Cloud.

Key Takeaways

• Five providers list Hindi, Spanish and Arabic together.
• Choose Arabic and Spanish voices by dialect, not by language name.
• A language list is not voice quality, so listen on a real call.

A provider listing a language does not mean its voice suits your callers on a phone line. Below, we map which voices cover Hindi, Spanish and Arabic, pick between them by script and dialect, give each caller a voice in their language and test every language before it goes live.

Which providers list Hindi, Spanish and Arabic

Several voice providers cover all three languages, and the difference is in which regional versions each one offers. Every provider in the table below runs in Dograh.

ElevenLabs' language list gives the same three for both models, and Flash v2.5 claims about 100 ms of latency, which matters on a live call. Cartesia's Sonic 3.6 page lists 44 languages.

Google's Chirp 3 HD voices cover Hindi and two Spanish locales, with one general Arabic locale. Read a "Yes" in the table as "the provider offers it", nothing more. How each voice sounds to your callers is a separate question, covered in the last section.

OpenAI's voices are left out. OpenAI says they are optimized for English, and its three text-to-speech models leave its API on 6 January 2027, as our post on OpenAI text-to-speech pricing explains.

Hindi and Hinglish on the same call

Many Hindi-speaking callers switch between Hindi and English inside one sentence, and the voice has to read that mix without stumbling. The first choice is the script. A language model asked to reply in Hinglish often writes it in Latin letters, while a voice trained on Hindi expects Devanagari.

Two providers address this directly. Cartesia's page notes expanded support for Hindi written in Latin script. Sarvam's model list gives English-Hindi as a language of its own on Bulbul v4 Flash, next to plain Hindi. Azure Speech has nine standard Hindi voices, plus newer HD voices.

Tell the language model which script to write in, and test that your chosen voice reads it well. Put the rule in the prompt, for example "Reply in Hindi written in Devanagari, and keep English product names in English." Our post on Hindi AI voice agents for collections covers Hindi and Hinglish speech on a full call, including fixed wording.

Numbers need the same decision. Many callers say amounts and dates in English even inside a Hindi sentence, while others expect Hindi numbers. Write the prompt to match how your callers speak, and test a price, a date and a phone number with the voice before launch.

Spanish from Spain, Mexico or the US

Spanish callers notice an accent from the wrong country, so pick the locale before the voice. Spain's Spanish and Mexico's sound different enough that a caller in Monterrey can tell the agent is not local.

The providers split this differently. ElevenLabs and Murf offer Spain and Mexico. Google's Chirp 3 HD offers Spain and US Spanish. Azure Speech covers all three, with HD and multilingual voices for Spain and Mexico and two standard voices for US Spanish. Deepgram's Aura-2 has 17 Spanish voices, and Spanish is the only one of the three languages it covers.

If your callers come from one country, match its locale. If they come from many, a neutral Latin American voice is a common middle ground, but test it with callers from your biggest markets before you commit.

Open Source Alternative to Vapi / Retell

Self-hosted voice agent platform — no per-minute fees

dograh-hq/dograh

Star on GitHub

Arabic is many dialects under one name

Arabic needs the most care, because a caller in Cairo and a caller in Riyadh speak noticeably different Arabic. A voice built on one dialect can sound foreign to the other, even when both read the same text.

Azure Speech's language list has the widest spread among these providers, with 16 Arabic locales such as the UAE, Saudi Arabia and Egypt. Each locale has two standard voices, and none of them is an HD voice. ElevenLabs lists Saudi and UAE Arabic. Google's Chirp 3 HD and Murf list one general Arabic locale, ar-XA.

So start with where your callers are. A Gulf bank wants a Gulf Arabic voice, which narrows the list to Azure's Gulf locales or ElevenLabs. A service line across North Africa and the Levant may need more than one voice.

The written text matters as much as the voice. Tell the language model which Arabic to write, Modern Standard Arabic or the caller's own dialect, and test that the voice reads it naturally. A Gulf voice reading formal written Arabic can still sound stiff to a caller who expects everyday speech. Our post on Arabic AI voice agents for banking covers Gulf dialects and keeping data in the UAE.

Giving each caller the right voice in Dograh

A voice is a setting, so serving three languages comes down to choosing the right voice for each call. Dograh is the orchestration layer that runs the call flow, tools and telephony, and the speech and language models are a separate layer you choose, on your own keys. Which languages an agent can speak, more than 70 in all, depends on the speech-to-text and voice providers you pick, which is why the provider table comes first.

Each model configuration uses one voice for every reply. In the voice settings, Google, Azure, Sarvam, Cartesia and Murf take a language or locale, while ElevenLabs takes its language from the text it is given. The simplest setup is one agent per language, each on its own number or campaign.

For outbound calls, the API Trigger can also choose the voice for one call when your system starts it, so a single agent can call a Spanish list and a Hindi list with the right voice for each.

When you already know a caller's language from your records, one agent can serve all three. Pre-Call Data Fetch asks your system about the caller once the call connects, and its answer can include a voice override for that call. A caller marked as Spanish then hears a Spanish voice from the first reply. The Greeting Text plays before Dograh can hear the caller, so write a short greeting in both languages, for example "Namaste, hello".

Speech-to-text is the other half. Our guide to multilingual voice agents covers which speech-to-text models switch languages mid-call. Fixed lines can skip the voice model altogether. Pre-recorded clips recorded by a native speaker in each language play the greeting and notices word for word.

When the agent is unsure which language it heard, it should not carry on guessing. Write the prompt so it asks a short question offering the likely languages and switches only after the caller confirms. If the caller's language is one the agent cannot serve well, hand the call to a person who speaks it with the Transfer Call tool, which works on Twilio, Telnyx, Plivo, Cloudonix and Asterisk ARI.

Test each language on a real phone line

A language list tells you a voice exists, and only listening tells you whether it works for your callers. Test every language as its own part of the agent, with a native speaker on the other end.

Test over the phone, not only in the browser. Most phone calls still reach the agent as 8 kHz audio, while a browser demo runs at 16 kHz, so a voice that sounds clear in a demo can sound flat on a call. Google's Chirp 3 HD streams in the phone formats mu-law and A-law, along with PCM.

Give each voice the lines that break voices: prices, dates, phone numbers, names and your own brand words. Ask native speakers whether the accent fits and whether the numbers sound right. Run the same short script in every language, about ten calls each, so the results compare like for like. Change one thing at a time, the voice, the locale or the prompt's script rule, and listen again before you change the next. Dograh's QA summaries are in English by default, so a reviewer who speaks the language still needs to listen to some calls in it. For what each voice costs per call, see our comparison of the best text-to-speech APIs for voice agents.