urdu-realtime-1: a realtime speech model for conversational Urdu

neu is training urdu-realtime-1, a speech-to-speech model for natural conversation in Urdu. The model replies in less than 200 ms.

Four sample calls in Urdu with urdu-realtime-1.

More than 200 million people in South Asia use Urdu and regional languages every day. Many voice tools can speak Urdu, but their speech does not sound like a natural conversation.

neu is training a new model for this purpose: urdu-realtime-1. Tashfeen, the founder of neu, published the first sample calls on X on 3 November 2026.

Sample calls

The video shows four short calls between a person and the model:

  • Food order: the caller orders chicken karahi without tomatoes.
  • Order confirmation: a shop confirms an order for shoes and gives the delivery time.
  • Business idea: the caller discusses a business idea for a car that uses water.
  • Household tasks: the caller and the model discuss tasks in the house.

In the order confirmation call, the model says آپ کے جوتوں کی قیمت دو ہزار روپے ہے ("The price of your shoes is 2,000 rupees"). Then the model asks for approval to send the order. It tells the caller that the order arrives in two to three days.

Model design

urdu-realtime-1 is a full-duplex speech-to-speech model. It listens and speaks at the same time, as in a telephone call. The model does not change speech into text before it replies. Thus, the response time is less than 200 ms.

The base model is an open full-duplex speech model from the Moshi family. We added a tokenizer for Urdu script. Then we trained the model in stages on these data:

  • Urdu podcasts with two speakers and checked transcripts
  • Recorded telephone calls
  • Urdu text

Most of the training occurs on our own DGX Spark cluster.

Conversation quality

Natural speech has rhythm and pauses. The training gives priority to these properties. During development, we found and corrected these problems:

  • Early versions of the model were frequently silent. The cause was the podcast data, in which one person speaks for a long time.
  • One version spoke at the same time as the caller. Now, a turn controller sets when the model speaks.
  • English words in Latin script caused errors in the text of the model. Now, the model writes only in Urdu script.

Current status

The model is in development. After the sample calls, we trained the model on more telephone calls. We also used reinforcement learning to make sure that the model answers the question of the caller.

If you speak Urdu, watch the video and send your feedback on X. To discuss an Urdu voice agent for your team, book a call.

Thank you for your time.

More posts

Need help with a project?

Discuss your task with us.

Talk to us