app.py

import os
import time
import librosa
import soundfile as sf
import numpy as np
import torch
import gradio as gr
from textwrap import wrap
from huggingface_hub import snapshot_download

# ✅ Trust XTTS classes
from torch.serialization import add_safe_globals
from TTS.config.shared_configs import BaseDatasetConfig
from TTS.tts.configs.xtts_config import XttsConfig, XttsAudioConfig, XttsArgs
add_safe_globals([XttsConfig, XttsAudioConfig, XttsArgs, BaseDatasetConfig])

from TTS.tts.models.xtts import Xtts

# ⏳ Global variables
model = None
config = None
gpt_cond = None
speaker_emb = None

# 🔁 Initialize model
def initialize_model():
    global model, config
    model_dir = snapshot_download("coqui/XTTS-v2")
    config = XttsConfig()
    config.load_json(f"{model_dir}/config.json")
    model = Xtts.init_from_config(config)
    model.load_checkpoint(config, checkpoint_dir=model_dir, use_deepspeed=False, eval=True)
    model.to(torch.device("cuda" if torch.cuda.is_available() else "cpu"))

# 🧠 Process speaker voice
def process_speaker_audio(file=None):
    global gpt_cond, speaker_emb
    if file:
        file_path = file.name
    elif os.path.exists("speaker.wav"):
        file_path = "speaker.wav"
    else:
        return "❌ No speaker file found."

    y, sr = librosa.load(file_path, sr=22050, mono=True)
    sf.write("speaker_cleaned.wav", y, samplerate=22050)
    gpt_cond, speaker_emb = model.get_conditioning_latents(audio_path=["speaker_cleaned.wav"])
    torch.save(gpt_cond, "gpt_cond.pt")
    torch.save(speaker_emb, "speaker_emb.pt")
    return "✅ Speaker voice processed!"

# 🔊 TTS Function
def synthesize(text, speed, temperature, top_k, top_p, speaker_file):
    global gpt_cond, speaker_emb
    if not text.strip():
        return "⚠️ Please enter some Hindi text.", None

    if gpt_cond is None or speaker_emb is None or speaker_file is not None:
        status = process_speaker_audio(speaker_file)
        if "❌" in status:
            return status, None

    chunks = wrap(text, 140)
    final_audio = []
    for chunk in chunks:
        out = model.inference(
            text=chunk,
            language="hi",
            gpt_cond_latent=gpt_cond,
            speaker_embedding=speaker_emb,
            speed=speed,
            temperature=temperature,
            length_penalty=1.0,
            repetition_penalty=2.0,
            top_k=top_k,
            top_p=top_p,
            do_sample=True,
        )
        final_audio.append(out["wav"])
    combined = np.concatenate(final_audio)
    output_path = "output.wav"
    sf.write(output_path, combined, samplerate=config.audio.output_sample_rate)
    return "🎉 Audio generated!", output_path

# 🚀 Launch Gradio App
def launch():
    gr.Interface(
        fn=synthesize,
        inputs=[
            gr.Textbox(lines=4, label="📝 Hindi Text"),
            gr.Slider(0.5, 1.5, value=0.9, step=0.05, label="🌀 Speed"),
            gr.Slider(0.1, 1.5, value=0.85, step=0.05, label="🔥 Temperature"),
            gr.Slider(10, 100, value=50, step=1, label="🎯 Top-K"),
            gr.Slider(0.1, 1.0, value=0.9, step=0.05, label="🎲 Top-P"),
            gr.File(label="🎤 Upload speaker.wav (optional)")
        ],
        outputs=[
            gr.Textbox(label="📣 Status"),
            gr.Audio(label="🔊 Output Audio")
        ],
        title="🇮🇳 XTTS Hindi Text-to-Speech (Custom Voice)",
        description="Enter Hindi text and customize speech generation. Upload a speaker file or use a default one in the repo.",
    ).launch()

initialize_model()
if os.path.exists("gpt_cond.pt") and os.path.exists("speaker_emb.pt"):
    gpt_cond = torch.load("gpt_cond.pt")
    speaker_emb = torch.load("speaker_emb.pt")

launch()