app.py

import nltk
nltk.download('all')
nltk.download('averaged_perceptron_tagger')
nltk.download('punkt')
import os
import uuid
import time
import torch
import gradio as gr
os.environ["NUMBA_DISABLE_CACHE"] = "1"
# import mecab_patch
# import english_patch
#from melo.api import TTS
from MeloTTS.melo.api import TTS
from openvoice.api import ToneColorConverter
#from meloTTS import english

# 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."

    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()))

    for speaker_key in speaker_ids.keys():
        speaker_id = speaker_ids[speaker_key]
        speaker_key = speaker_key.lower().replace('_', '-')

    # Generate base TTS voice
        speed = 1.0
        #source_se = torch.load(f'checkpoints/base_speakers/EN/{speaker_key}.pth', map_location=device)
        model.tts_to_file(text, speaker_id, tmp_melo_path,speed=speed)

    # Use speaker_wav as reference to extract style embedding
        from openvoice import se_extractor
        torch.hub.load('snakers4/silero-vad', 'silero_vad', force_reload=False)
        ref_se, _ = se_extractor.get_se(speaker_wav, tone_color_converter, vad=True)

    # 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

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()

# iface = gr.Interface(
#     fn=clone_with_base_speaker,
#     inputs=[
#         gr.Textbox(label="Input Text", placeholder="Enter text to synthesize..."),
#         gr.Dropdown(choices=base_speaker_choices, label="Select Base Speaker"),
#     ],
#     outputs=gr.Audio(type="filepath", label="Cloned Voice Output"),
#     title="Voice Cloning with OpenVoice Base Speakers",
#     description="Choose a base speaker from OpenVoice and enter text to generate voice."
# )

# iface.launch()


# import os
# import time
# import uuid
# import gradio as gr

# from TTS.api import TTS
# from openvoice import se_extractor
# from openvoice.api import ToneColorConverter

# # Import your local english.py logic
# from meloTTS import english

# # Paths
# device = "cuda" if os.system("nvidia-smi") == 0 else "cpu"
# output_dir = "outputs"
# os.makedirs(output_dir, exist_ok=True)

# # Load OpenVoice tone converter
# tone_color_converter = ToneColorConverter(f"{os.getcwd()}/checkpoints", device=device)
# tone_color_converter.load_model()

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

#     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
#     model.tts_to_file(text, speaker_id=default_speaker_id, file_path=tmp_melo_path, speed=1.0)

#     # Extract style embedding
#     ref_se, _ = se_extractor.get_se(speaker_wav, tone_color_converter, vad=False)

#     # Convert tone
#     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
# demo = gr.Interface(
#     fn=clone_and_speak,
#     inputs=[
#         gr.Textbox(label="Text to Synthesize"),
#         gr.Audio(label="Reference Voice (WAV)", type="filepath")
#     ],
#     outputs=gr.Audio(label="Cloned Voice Output"),
#     title="Voice Cloner with MeloTTS + OpenVoice"
# )

# if __name__ == "__main__":
#     demo.launch()