app.py

from transformers import pipeline, AutoTokenizer, AutoModelForTokenClassification
import gradio as gr
import numpy as np
import scipy.io.wavfile
import tempfile
import os
from transformers import VitsModel, AutoTokenizer
import torch
import re
import traceback

print("Starting application...")

# Global variables for models
punct_pipe = None
model = None
tokenizer = None

def load_models():
    global punct_pipe, model, tokenizer
    
    print("Loading punctuation model...")
    try:
        punctuation_model_id = "oliverguhr/fullstop-punctuation-multilang-large"
        punct_tokenizer = AutoTokenizer.from_pretrained(punctuation_model_id)
        punct_model = AutoModelForTokenClassification.from_pretrained(punctuation_model_id)
        punct_pipe = pipeline("token-classification", model=punct_model, tokenizer=punct_tokenizer, aggregation_strategy="simple")
        print("✓ Punctuation model loaded successfully")
    except Exception as e:
        print(f"✗ Error loading punctuation model: {e}")
        punct_pipe = None

    print("Loading TTS model...")
    try:
        model = VitsModel.from_pretrained("facebook/mms-tts-kmr-script_latin")
        tokenizer = AutoTokenizer.from_pretrained("facebook/mms-tts-kmr-script_latin")
        print("✓ TTS model loaded successfully")
    except Exception as e:
        print(f"✗ Error loading TTS model: {e}")
        model = None
        tokenizer = None

# Load models at startup
load_models()

# Simple number-to-Kurmanji-word mapping
num2word = {
    "0": "sifir", "1": "yek", "2": "du", "3": "sê", "4": "çar", "5": "pênc",
    "6": "şeş", "7": "heft", "8": "heşt", "9": "neh", "10": "deh"
}

def replace_numbers_with_words(text):
    def repl(match):
        num = match.group()
        return num2word.get(num, num)
    return re.sub(r'\b\d+\b', repl, text)

def restore_punctuation(text):
    if punct_pipe is None:
        print("Punctuation model not available, skipping...")
        return text
    
    try:
        results = punct_pipe(text)
        punctuated = ""
        for token in results:
            word = token['word']
            punct = token.get('entity_group', '')
            if punct == "PERIOD":
                punctuated += word + ". "
            elif punct == "COMMA":
                punctuated += word + ", "
            else:
                punctuated += word + " "
        return punctuated.strip()
    except Exception as e:
        print(f"Punctuation error: {e}")
        return text

def text_to_speech(text):
    print(f"=== TTS Function Called ===")
    print(f"Input text: '{text}'")
    
    try:
        # Basic validation
        if not text or text.strip() == "":
            error_msg = "Please enter some text"
            print(f"Error: {error_msg}")
            return None
        
        # Check if models are loaded
        if model is None or tokenizer is None:
            error_msg = "TTS model not loaded properly"
            print(f"Error: {error_msg}")
            return None
        
        print("Processing text...")
        
        # Process text
        processed_text = text.strip()  # Start simple, skip punctuation for now
        processed_text = replace_numbers_with_words(processed_text)
        print(f"Processed text: '{processed_text}'")
        
        # Tokenize
        print("Tokenizing...")
        inputs = tokenizer(processed_text, return_tensors="pt")
        print(f"Tokenized successfully, input_ids shape: {inputs['input_ids'].shape}")
        
        # Generate audio
        print("Generating audio...")
        with torch.no_grad():
            output = model(**inputs).waveform
        print(f"Audio generated, shape: {output.shape}")
        
        # Convert to numpy
        waveform = output.squeeze().numpy()
        print(f"Waveform shape: {waveform.shape}")
        
        # Save to file
        print("Saving audio file...")
        tmp_file = tempfile.NamedTemporaryFile(suffix=".wav", delete=False)
        tmp_path = tmp_file.name
        tmp_file.close()
        
        scipy.io.wavfile.write(
            tmp_path, 
            rate=model.config.sampling_rate, 
            data=waveform
        )
        
        print(f"✓ Audio saved to: {tmp_path}")
        print("=== TTS Function Completed Successfully ===")
        return tmp_path
        
    except Exception as e:
        error_msg = f"Error in TTS: {str(e)}"
        print(f"✗ {error_msg}")
        print("Full traceback:")
        traceback.print_exc()
        return None

# Simple test function to verify Gradio is working
def test_function(text):
    print(f"Test function called with: {text}")
    return f"You entered: {text}"

# Create a simple interface first to test
print("Creating Gradio interface...")

# Option 1: Simple test interface (uncomment to test basic functionality)
# interface = gr.Interface(
#     fn=test_function,
#     inputs=gr.Textbox(label="Test Input"),
#     outputs=gr.Textbox(label="Test Output"),
#     title="Test Interface"
# )

# Option 2: Full TTS interface
interface = gr.Interface(
    fn=text_to_speech,
    inputs=gr.Textbox(
        label="Enter Kurmanji Text",
        placeholder="e.g. Silav! Ez bi xêr im.",
        lines=2,
        value=""  # Default empty value
    ),
    outputs=gr.Audio(label="Generated Speech"),
    title="Kurmanji Text-to-Speech",
    description="Enter Kurmanji Kurdish text to convert to speech.",
    examples=[
        ["Silav"],
        ["Ez bi xêr im"],
        ["Spas"]
    ],
    cache_examples=False,
    flagging_mode="never"
)

print("Launching interface...")

if __name__ == "__main__":
    interface.launch(
        debug=True,
        share=False,
        show_error=True,
        server_name="0.0.0.0" if "SPACE_ID" in os.environ else "127.0.0.1",
        server_port=7860
    )