README.md

metadata

title: Parler TTS API
emoji: 🎙️
colorFrom: blue
colorTo: green
sdk: docker
app_file: api.py
python_version: 3.1

Indic Parler-TTS API

FastAPI endpoint for Urdu Text-to-Speech using ai4bharat/indic-parler-tts.

API Endpoints

Health Check

GET /

Returns model status and available speakers.

Response:

{
  "status": "ok",
  "model": "Indic Parler-TTS",
  "speakers": ["Divya", "Rani", "Rohit", "Aman", "Generic Female", "Generic Male"],
  "sample_rate": 24000
}

Generate Speech

POST /tts

Request Body:

{
  "text": "السلام علیکم، میرا نام اردو ٹی ٹی ایس ہے۔",
  "speaker": "Divya",
  "pitch": "Moderate",
  "rate": "Moderate",
  "temperature": 0.8,
  "do_sample": true
}

Parameters:

• text (string, required): Urdu text to synthesize
• speaker (string, optional): Speaker name. Options: Divya, Rani, Rohit, Aman, Generic Female, Generic Male. Default: Divya
• pitch (string, optional): Voice pitch. Options: High, Moderate, Low. Default: Moderate
• rate (string, optional): Speaking rate. Options: Slow, Moderate, Fast. Default: Moderate
• temperature (float, optional): Sampling temperature (0.1-2.0). Default: 0.8
• do_sample (boolean, optional): Use sampling vs greedy decoding. Default: true

Response:

• WAV audio file (audio/wav)

Get Available Speakers

GET /speakers

Response:

{
  "speakers": ["Divya", "Rani", "Rohit", "Aman", "Generic Female", "Generic Male"]
}

Example Usage

cURL

curl -X POST http://localhost:7860/tts \
  -H "Content-Type: application/json" \
  -d '{
    "text": "السلام علیکم",
    "speaker": "Divya",
    "pitch": "Moderate",
    "rate": "Moderate"
  }' \
  --output speech.wav

Python

import requests
import json

url = "http://localhost:7860/tts"
payload = {
    "text": "السلام علیکم، میرا نام اردو ٹی ٹی ایس ہے۔",
    "speaker": "Divya",
    "pitch": "Moderate",
    "rate": "Moderate",
    "temperature": 0.8,
    "do_sample": True
}

response = requests.post(url, json=payload)
if response.status_code == 200:
    with open("speech.wav", "wb") as f:
        f.write(response.content)
    print("Audio saved!")
else:
    print(f"Error: {response.status_code}")
    print(response.text)

Running Locally

With Docker

docker build -t parler-tts-api .
docker run -p 7860:7860 --gpus all parler-tts-api

Without Docker

python3 -m venv venv
source venv/bin/activate
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
pip install -r requirements.txt
pip install uvicorn[standard]
python api.py

Then visit http://localhost:7860/docs for interactive API documentation.

Environment Variables

For HF Spaces deployment, set the following secret:

• HF_TOKEN: Your Hugging Face API token (required for gated model access)

Technical Details

• Model: Indic Parler-TTS (multi-speaker, multi-language)
• Language: Urdu (auto-detected from script)
• Sample Rate: 24 kHz
• Audio Format: WAV (16-bit PCM)
• Framework: FastAPI + PyTorch
• Deployment: HF Spaces Docker runtime

Quality Notes

• Language is auto-detected from Urdu script — do NOT mention language in voice descriptions
• Named speakers (Divya, Rohit, etc.) provide consistent voices
• Same random seed used across sentences for voice consistency within a generation
• Text cleaning removes Latin/English characters to prevent language mixing