app.py

import torch
from transformers import SpeechT5ForTextToSpeech, SpeechT5Processor
import soundfile as sf
import gradio as gr

# Load the TTS model from the Hugging Face Hub
checkpoint = "arham061/speecht5_finetuned_voxpopuli_nl"  # Replace with your actual model name
processor = SpeechT5Processor.from_pretrained(checkpoint)
model = SpeechT5ForTextToSpeech.from_pretrained(checkpoint)
tokenizer = processor.tokenizer


# Buckwalter to Unicode mapping
buck2uni = {
    u"\u0627": "A",
    u"\u0675": "A",
    u"\u0673": "A",
    u"\u0630": "A",
    u"\u0622": "AA",
    # Rest of the mapping...
}

def transString(string, reverse=0):
    """Given a Unicode string, transliterate into Buckwalter. To go from
    Buckwalter back to Unicode, set reverse=1"""
    for k, v in buck2uni.items():
        if not reverse:
            string = string.replace(k, v)
        else:
            string = string.replace(v, k)
    return string


def generate_audio(text):
    # Convert input text to Roman Urdu
    roman_urdu = transString(text)

    # Tokenize the input text
    inputs = tokenizer(roman_urdu, return_tensors="pt").input_values

    # Generate speech from the model
    with torch.no_grad():
        logits = model(inputs).logits

    # Convert logits to audio waveform
    predicted_ids = torch.argmax(logits, dim=-1)
    audio = tokenizer.decode(predicted_ids[0], skip_special_tokens=True)

    return audio


def text_to_speech(text):
    # Generate audio
    audio_output = generate_audio(text)

    # Save audio as a .wav file
    sf.write("output.wav", audio_output, samplerate=22050)

    return "output.wav"


# Define the Gradio interface
inputs = gr.inputs.Textbox(label="Enter text in Urdu")
outputs = gr.outputs.Audio(label="Audio")

interface = gr.Interface(fn=text_to_speech, inputs=inputs, outputs=outputs, title="Urdu TTS")
interface.launch()