app.py

import os
import uuid
import time
import torch
import gradio as gr
os.environ["NUMBA_DISABLE_CACHE"] = "1"
import mecab_patch
from melo.api import TTS
from openvoice.api import ToneColorConverter

# Set temporary cache locations for Hugging Face Spaces
os.environ["TORCH_HOME"] = "/tmp/torch"
os.environ["HF_HOME"] = "/tmp/huggingface"
os.environ["HF_HUB_CACHE"] = "/tmp/huggingface"
os.environ["TRANSFORMERS_CACHE"] = "/tmp/huggingface"
os.environ["MPLCONFIGDIR"] = "/tmp"
os.environ["XDG_CACHE_HOME"] = "/tmp"
os.environ["XDG_CONFIG_HOME"] = "/tmp"
os.environ["NUMBA_DISABLE_CACHE"] = "1"
os.makedirs("/tmp/torch", exist_ok=True)
os.makedirs("/tmp/huggingface", exist_ok=True)
os.makedirs("/tmp/flagged", exist_ok=True)

# Output folder
output_dir = "/tmp/outputs"
os.makedirs(output_dir, exist_ok=True)

# Initialize tone converter
ckpt_converter = "checkpoints/converter/config.json"
tone_color_converter = ToneColorConverter(ckpt_converter)

# Device setting
device = "cuda" if torch.cuda.is_available() else "cpu"

def clone_and_speak(text, speaker_wav):
    if not speaker_wav:
        return "Please upload a reference .wav file."

    # import melo.text.english as english
    # original_g2p = english.g2p

    # def patched_g2p(text):
    #     phones, tones, word2ph = original_g2p(text)
    #     # Fix: wrap ints in list to avoid TypeError
    #     word2ph_fixed = []
    #     for item in word2ph:
    #         if isinstance(item, int):
    #             word2ph_fixed.append([item])
    #         else:
    #             word2ph_fixed.append(item)
    #     return phones, tones, word2ph_fixed

    # english.g2p = patched_g2p

    base_name = f"output_{int(time.time())}_{uuid.uuid4().hex[:6]}"
    tmp_melo_path = f"{output_dir}/{base_name}_tmp.wav"
    final_output_path = f"{output_dir}/{base_name}_converted.wav"

    # Use English speaker model
    model = TTS(language="EN", device=device)
    speaker_ids = model.hps.data.spk2id
    default_speaker_id = next(iter(speaker_ids.values()))

    # Generate base TTS voice
    speed = 1.0
    import english_patch
    model.tts_to_file(text, default_speaker_id, tmp_melo_path,speed=speed)

    # Use speaker_wav as reference to extract style embedding
    from openvoice import se_extractor
    ref_se, _ = se_extractor.get_se(speaker_wav, tone_color_converter, vad=False)

    # Run the tone conversion
    tone_color_converter.convert(
        audio_src_path=tmp_melo_path,
        src_se=ref_se,
        tgt_se=ref_se,
        output_path=final_output_path,
        message="@HuggingFace",
    )

    return final_output_path

# Gradio interface
gr.Interface(
    fn=clone_and_speak,
    inputs=[
        gr.Textbox(label="Enter Text"),
        gr.Audio(type="filepath", label="Upload a Reference Voice (.wav)")
    ],
    outputs=gr.Audio(label="Synthesized Output"),
    flagging_dir="/tmp/flagged",
    title="Text to Voice using Melo TTS + OpenVoice",
    description="Use Melo TTS for base synthesis and OpenVoice to apply a reference speaker's tone.",
).launch()