Urdu Voice Agents: Can AI Handle Urdu and Roman Urdu?

Most pages on this topic say "Urdu supported" and stop. Here is what works on a live line, where it breaks, and how to test it.

Can AI speak Urdu on phone calls?

Yes, well enough for focused jobs: booking appointments, taking orders, answering fee and timing questions, routing calls. It is not ready to run unsupervised on long, open-ended calls with elderly callers on noisy lines.

A voice agent is a loop of three parts: speech-to-text (STT) turns the caller's voice into text, a language model decides the reply, and text-to-speech (TTS) speaks it. English has strong options at every step. Urdu has a weak link at each end. The language model in the middle is the easy part, since current models handle Urdu script and Roman Urdu well.

So pick your Urdu STT and TTS first, then build around them. Otherwise you get an agent that sounds fine in a demo and fails on real calls.

Which speech-to-text works best for Urdu, including Urdu-English mixing?

Real callers rarely speak "pure" Urdu. A typical sentence is "Mujhe kal ka appointment chahiye, cleaning ke liye": Urdu grammar, English nouns. STT models that expect one language per stream either force the English words into Urdu spelling or drop them.

What to look for:

• In-stream language switching. Soniox advertises real-time Urdu STT that switches languages inside one stream, across 60+ languages, from $0.12 per hour of audio (as of October 2026). That is about $0.002 a minute, a small share of your bill.
• Platform support. Retell lists Urdu among 55+ supported voice agent languages. That does not tell you which STT engine runs underneath, so ask.
• Output script. Engines return Urdu in Nastaliq (اردو), in Roman Urdu, or mixed. Your prompts must expect the script the engine actually returns.

We would not name a "best" Urdu STT without testing on your callers. Accents differ between Karachi, Lahore, Peshawar and the diaspora, and phone audio is compressed. Run your own recordings through two or three engines and compare.

Does ElevenLabs support Urdu, and which voice models do?

ElevenLabs' language documentation lists Urdu (ur) under Eleven v3 only. It is not in the Multilingual v2 list (29 languages) or the Flash v2.5 list (32 languages) as of October 2026. Flash is what most phone agents use because it responds fast, so test v3's response time on a real call, not a browser preview.

Three practical points on Urdu TTS:

1. Feed it Urdu script. A voice reading "kal" or "mein" in Latin letters may pronounce it like English. Generate spoken replies in Urdu script and keep Roman Urdu for text channels.
2. Write numbers out. Decide whether "Rs. 2,500" is spoken in Urdu or English, and convert it to words before it reaches the voice.
3. Get a native speaker to listen. A voice that sounds foreign makes callers switch to English or hang up.

Can a chatbot reply in Roman Urdu on WhatsApp?

Yes, and this is where Urdu agents work best today. Most people type Roman Urdu on WhatsApp ("kal ka time mil sakta hai?"), and language models reply in it comfortably. With no STT or TTS in the loop, two of the three weak links disappear.

• Mirror the customer. Reply in whichever script they wrote in.
• Expect spelling variety. "Hai", "hy" and "he" mean the same thing, as do "kya" and "kia". Never build logic on exact keyword matches.
• Keep messages short. Two lines and a clear next step.

For many businesses the right order is a Roman Urdu WhatsApp agent first, then voice once the booking logic is proven. See our guide to setting up a WhatsApp AI agent for your business.

Why are names, numbers and addresses the hard part?

Getting the intent right is easy. Getting the details right is where Urdu calls fail.

• Names. "Ayesha", "Aisha" and "Aaisha" are one person. Match against records loosely and confirm.
• Phone numbers. Callers mix languages inside one number: "zero three hundred, ek do teen, four five six seven." Read every number back digit by digit, or let callers use the keypad.
• Money. "Dhai hazaar", "paune do lakh" and "2.5k" all appear. Normalize to digits and say the figure back.
• Addresses. Sector addresses like "G-11/2, street 5" are easy to garble. Ask for the address by WhatsApp, typed or pinned, instead of by voice.

A well-designed dental clinic booking sounds like this:

The agent offers two slots instead of an open question, reads the number back, and moves written details to WhatsApp. Each choice removes a place where recognition can go wrong.

How do you test an Urdu agent before go-live?

1. Collect real audio. Record (with consent) 30 to 50 typical calls, or have staff role-play them the way customers really talk.
2. Cover accents and conditions. Punjabi, Pashto and Sindhi-influenced Urdu, older callers, speakerphone, street noise.
3. Score the fields. Check whether name, number, date, time and service landed correctly in the calendar or CRM.
4. Test the handover. Say "mujhe kisi insaan se baat karni hai" (I want to talk to a person) and confirm it transfers with a summary attached.
5. Go live narrow. Start with after-hours calls, review every transcript for two weeks, then widen.

Is an Urdu voice agent more expensive than English?

Usually, because of the voice. Retell lists ElevenLabs v3 TTS at $0.10 per minute against $0.015 for standard voices, with voice infrastructure at $0.055 and telephony around $0.015 per minute (as of October 2026). Before the language model:

The language model adds a few cents per minute to both, and Urdu text can use more tokens than English for the same sentence. For the full breakdown across Vapi, Retell, ElevenLabs and Twilio, see how much an AI voice agent costs in 2026. An extra $85 per 1,000 minutes is small next to a salary, but check your call mix before paying for Urdu "just in case."

FAQ

Can AI speak Urdu on phone calls?

Yes. Platforms such as Retell list Urdu as a supported language, and ElevenLabs offers Urdu voices on Eleven v3. Urdu agents work well for focused tasks like bookings, orders and FAQs. They need extra care with names, numbers and addresses, and should hand over to a person when a caller asks or the agent is unsure.

Does ElevenLabs support Urdu?

Yes, but only on one model. ElevenLabs' language documentation lists Urdu under Eleven v3. It is not listed for Multilingual v2 (29 languages) or Flash v2.5 (32 languages) as of October 2026. Because v3 is the only option, test response time on a real phone call and check the per-minute price on your voice platform before committing.

Can a voice agent understand callers who mix Urdu and English?

It can if the speech-to-text engine handles language switching inside one audio stream. Soniox, for example, advertises in-stream switching across 60+ languages including Urdu. Engines that expect one language per call mangle English words inside Urdu sentences. Test with real recordings, because most Pakistani callers code-switch constantly.

Can a chatbot reply in Roman Urdu?

Yes. Current language models read and write Roman Urdu well, which makes WhatsApp the easiest place to start an Urdu agent. Set it to mirror the customer's script, expect spellings like "kya" and "kia", and keep replies short. With no speech recognition or voice in the loop, chat agents avoid most Urdu-specific errors.

Is an Urdu voice agent more expensive than English?

Usually, because of the voice. On Retell's published rates as of October 2026, ElevenLabs v3 text-to-speech costs $0.10 per minute against $0.015 for standard voices. With infrastructure and telephony, that is about $0.17 versus $0.085 per minute before the language model, or roughly $170 versus $85 for 1,000 minutes a month.

Which speech-to-text works best for Urdu?

There is no universal winner. Look for real-time Urdu support with in-stream language switching, such as Soniox, which starts at $0.12 per hour of audio. Then compare two or three engines on 30 to 50 of your own recorded calls, scoring how often names, numbers and intents come out right.

Where to go from here

Start with the channel where your customers already talk to you, prove the logic, then add voice. Shorai Studios, an AI studio that builds voice and WhatsApp agents in English, Urdu and Roman Urdu, can scope that with you. See our AI voice and WhatsApp agent services, or look at the AI front desk we built for dental clinics, which books, reschedules and sends reminders over calls and WhatsApp.

We build the first version of your agent for free, so you can hear it handle your own callers before you pay anything.

Sources

• ElevenLabs: What languages do you support?
• Soniox: Urdu speech-to-text for voice agents
• Retell AI: Urdu AI voice agents
• Retell AI: Pricing