Urdu-speaking AI models — really really desi style

We have trained an AI model called Orator, that speaks Urdu with human-like realism. To the best of our knowledge, it is the first model that can speak Urdu in an authentic way: the way everyday people speak Urdu in everyday life.

We have voices from “the whole mohalla“ (entire neighborhood). From angry molvis and nosey aunties, to thailay walas, and burger boys.

You can experience the voices at UpliftAI.org. The website will allow you to experience dozens of voices in seconds by just clicking/hovering the circles.

Making the model

We have put our heart into this work, and enjoying it more than any work we have ever done in the past.

We started with the map of Pakistan and asked: how do people sound in every district. That gave us the accent list.

Then, we asked ourselves: what are all the unique characters we would interact with in a typical month as we go about our normal life. This gave us our character/persona list.

We then hired a leading voice producer and worked with him to hire voice actors and script writers, and then spent months doing voice recordings in studios in Karachi, Lahore, Quetta and Islamabad.

Once the voices were recorded, then we hired 50 contractors to label data.

The final and the simplest step was the model training. We trained a very small 32M param VITS2 inspired model.

The results are very encouraging.

- Here is someone creating a story book using our voices: https://shehryar-stories.up.railway.app/frontend/index.html

- Here is me just having fun with it: https://www.linkedin.com/posts/hammad2_is-your-mom-complaining-about-massi-leaving-ugcPost-7468570160233107457-sbQ5/

Ask

• If you know anyone that may benefit from this, please do connect us.
• If you know any actors, dubbers or media houses that would be interested in recording more characters, please do connect us.
• I can be reached at: https://www.linkedin.com/in/hammad2/ or founders@upliftai.org