app.py

import torch
from parler_tts import ParlerTTSForConditionalGeneration
from transformers import AutoTokenizer
import gradio as gr
import numpy as np
import time
import requests

# Set device to GPU if available, else CPU
device = "cuda:0" if torch.cuda.is_available() else "cpu"

# Function to load the model with retry logic and increased timeout
def load_model_with_retry(model_id, retries=3, timeout=60):
    for attempt in range(1, retries + 1):
        try:
            model = ParlerTTSForConditionalGeneration.from_pretrained(model_id, timeout=timeout).to(device)
            return model
        except requests.exceptions.ReadTimeout:
            print(f"Timeout when loading model. Attempt {attempt} of {retries}. Retrying in 5 seconds...")
            time.sleep(5)
    raise Exception("Failed to load model after multiple retries.")

# Function to load a tokenizer with retry logic and increased timeout
def load_tokenizer_with_retry(tokenizer_id, retries=3, timeout=60):
    for attempt in range(1, retries + 1):
        try:
            tokenizer = AutoTokenizer.from_pretrained(tokenizer_id, timeout=timeout)
            return tokenizer
        except requests.exceptions.ReadTimeout:
            print(f"Timeout when loading tokenizer. Attempt {attempt} of {retries}. Retrying in 5 seconds...")
            time.sleep(5)
    raise Exception("Failed to load tokenizer after multiple retries.")

# Load the TTS model and tokenizers using the retry functions
model = load_model_with_retry("ai4bharat/indic-parler-tts")
tokenizer = load_tokenizer_with_retry("ai4bharat/indic-parler-tts")
description_tokenizer = load_tokenizer_with_retry(model.config.text_encoder._name_or_path)

def generate_audio(text: str):
    """
    Generate synthesized speech audio based on the input text.
    
    Args:
        text (str): The text prompt to be spoken.
    
    Returns:
        tuple: A tuple containing the audio numpy array and the sampling rate.
    """
    # Set a default voice description
    default_description = (
        "Divya's voice is monotone yet slightly fast in delivery, with a very close recording "
        "that almost has no background noise."
    )
    
    # Tokenize the default description and the input text
    description_tokens = description_tokenizer(default_description, return_tensors="pt").to(device)
    prompt_tokens = tokenizer(text, return_tensors="pt").to(device)
    
    # Generate the audio tensor using the model
    generation = model.generate(
        input_ids=description_tokens.input_ids, 
        attention_mask=description_tokens.attention_mask, 
        prompt_input_ids=prompt_tokens.input_ids, 
        prompt_attention_mask=prompt_tokens.attention_mask
    )
    
    # Convert the generated tensor to a numpy array and remove extra dimensions
    audio_arr = generation.cpu().numpy().squeeze()
    
    # Retrieve the sampling rate from the model config
    sampling_rate = model.config.sampling_rate
    return (audio_arr, sampling_rate)

# Build the Gradio interface with a single text input
iface = gr.Interface(
    fn=generate_audio,
    inputs=gr.Textbox(label="Enter Text", value="अरे, तुम आज कैसे हो?"),
    outputs=gr.Audio(label="Generated Audio"),
    title="Indic Parler TTS",
    description="Generate synthesized speech using the Indic Parler TTS model from ai4bharat."
)

if __name__ == "__main__":
    iface.launch()