app.py

# app.py
import streamlit as st
import json
import soundfile as sf
import io
from tts_utils import load_model, generate_speech

import torch
torch.backends.cudnn.benchmark = True  # Enable cuDNN benchmark for performance
torch.cuda.empty_cache()  # Clear CUDA cache to prevent memory issues

# Cache model loading for performance
@st.cache_resource
def cached_load_model():
    return load_model()

# Load voice database from JSON
def load_voice_database():
    with open('indian_voices.json') as f:
        voices = json.load(f)
    return {voice['name']: voice for voice in voices}

INDIAN_FEMALE_VOICES = load_voice_database()

# Streamlit UI
st.title("🎙️ Indian English TTS Voice Selector")
text_input = st.text_area("Enter your text:", "Hey, how are you doing today?")

# Voice selection with warnings
selected_voice = st.selectbox(
    "Choose voice:", 
    options=list(INDIAN_FEMALE_VOICES.keys()),
    format_func=lambda x: f"{x} 🔊"  # Add speaker icon
)

# Show voice description and warning
voice_info = INDIAN_FEMALE_VOICES[selected_voice]
st.caption(f"**Description:** {voice_info['description']}")
if voice_info['warning']:
    st.warning(f"⚠️ Note: {voice_info['warning']}")

if st.button("Generate Speech", type="primary"):
    model, tokenizer, desc_tokenizer = cached_load_model()
    
    with st.spinner(f"Generating {selected_voice}'s voice..."):
        try:
            audio_array = generate_speech(
                text_input,
                voice_info['prompt'],
                model,
                tokenizer,
                desc_tokenizer
            )
            
            # Audio playback
            buffer = io.BytesIO()
            sf.write(buffer, audio_array, 16000, format="WAV")
            buffer.seek(0)
            
            st.audio(buffer, format="audio/wav")
            st.success("Audio generated successfully!")
            
        except RuntimeError as e:
            st.error(f"GPU Error: {str(e)}. Try a different voice or shorter text.")