README.md

---
title: Urdu S2S MVP
colorFrom: green
colorTo: blue
sdk: gradio
sdk_version: 6.8.0
pinned: false
license: other
short_description: Urdu speech-to-speech and text-to-speech demo
---

# Urdu S2S MVP

This Space runs the current Urdu speech experience: a Pakistani Urdu voice assistant
demo for conversational speech-to-speech and text-to-speech testing.

Use **Speech to speech** to record or upload audio and receive a spoken Urdu response.
Use **Text to speech** to type Urdu, Roman Urdu, or Devanagari text and generate speech.

## About the Model

Urdu S2S is an experimental voice model focused on natural spoken Pakistani Urdu,
short assistant-style replies, and a warm female assistant voice. It is intended for
interactive product testing rather than long-form narration.

## Architecture

Urdu S2S uses a modular speech pipeline with dedicated components for speech recognition,
Urdu response generation, pronunciation-aware speech preparation, and neural speech
synthesis. The user-facing experience is a single speech-to-speech model demo, while the
underlying design lets each part improve independently as stronger Urdu speech models
become available.

At a high level, incoming audio is normalized and transcribed, the dialogue layer produces
a concise Urdu assistant response, a pronunciation layer prepares the response for stable
speech generation, and the synthesis layer renders the final spoken voice.

The current demo is tuned for:

- Urdu speech-to-speech conversations
- Urdu, Roman Urdu, and Devanagari text-to-speech
- Short support, scheduling, helpdesk, and everyday assistant prompts
- Urdu-first wording and pronunciation preferences
- Neutral user addressing without assuming the speaker's gender

## How to Test

For best results, use short recordings of 2 to 10 seconds with clear speech. Try prompts
such as greetings, clarification requests, simple account-support questions, scheduling
drafts, travel questions, or customer-service style requests.

## Current Limits

This is an MVP. Very noisy audio, long recordings, dense factual questions, unusual names,
and heavy code-switching may still create transcription or pronunciation errors. The first
generation after the Space wakes up may be slower while the speech model loads.

## Deployment

This Space is configured for interactive demo usage with GPU-backed speech generation.