app.py

import os
os.environ["NUMBA_DISABLE_CACHE"] = "1"

import gradio as gr
from docx import Document
from TTS.api import TTS
import tempfile

VOICE_MODELS = {
    "Jenny (Expressive Female)": {
        "model_name": "tts_models/en/jenny/jenny",
        "multi_speaker": False
    },
    "LJSpeech (Standard Female)": {
        "model_name": "tts_models/en/ljspeech/vits",
        "multi_speaker": False
    },
    "VCTK (Multiple Speakers)": {
        "model_name": "tts_models/en/vctk/vits",
        "multi_speaker": True
    }
}

MODEL_CACHE = {}

def load_tts_model(model_key):
    if model_key in MODEL_CACHE:
        return MODEL_CACHE[model_key]
    info = VOICE_MODELS[model_key]
    tts = TTS(model_name=info["model_name"], progress_bar=False, gpu=False)
    MODEL_CACHE[model_key] = tts
    return tts

def extract_speakers(model_key):
    info = VOICE_MODELS[model_key]
    if info["multi_speaker"]:
        if info["model_name"] == "tts_models/en/vctk/vits":
            return ["p225", "p226", "p227", "p228", "p229", "p230", "p231", "p232", "p233", "p234"]
        else:
            tts = load_tts_model(model_key)
            return getattr(tts, "speakers", [])
    return []

def docx_to_wav(doc_file, selected_voice, selected_speaker=None):
    info = VOICE_MODELS[selected_voice]
    tts = load_tts_model(selected_voice)

    document = Document(doc_file.name)
    full_text = "\n".join([para.text for para in document.paragraphs if para.text.strip()])

    with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tmp_wav:
        wav_path = tmp_wav.name

    kwargs = {}
    if info["multi_speaker"]:
        kwargs["speaker"] = selected_speaker

    tts.tts_to_file(text=full_text, file_path=wav_path, **kwargs)
    return wav_path

def show_load_button(voice_selection):
    info = VOICE_MODELS[voice_selection]
    return gr.update(visible=info["multi_speaker"]), gr.update(visible=False), gr.update(visible=False, interactive=False)

def load_and_show_speakers(voice_selection):
    speakers = extract_speakers(voice_selection)
    return gr.update(choices=speakers, visible=True, value=speakers[0]), gr.update(interactive=True)

with gr.Blocks() as interface:
    gr.Markdown("# 🎤 Realistic Voiceover from DOCX\nUpload a `.docx` file, select a voice, and generate lifelike speech!")

    with gr.Row():
        docx_input = gr.File(label="Upload .docx File", type="filepath")
        voice_dropdown = gr.Dropdown(choices=list(VOICE_MODELS.keys()), value="Jenny (Expressive Female)", label="Voice")
        load_speakers_btn = gr.Button("🔍 Load Speakers", visible=False)
        speaker_dropdown = gr.Dropdown(choices=[], label="Speaker", visible=False)
    
    generate_button = gr.Button("🎧 Generate Speech", interactive=True)
    audio_output = gr.Audio(label="🔊 Download WAV", type="filepath")

    # Interactions
    voice_dropdown.change(fn=show_load_button, inputs=voice_dropdown, outputs=[load_speakers_btn, speaker_dropdown, generate_button])
    load_speakers_btn.click(fn=load_and_show_speakers, inputs=voice_dropdown, outputs=[speaker_dropdown, generate_button])

    generate_button.click(fn=docx_to_wav, inputs=[docx_input, voice_dropdown, speaker_dropdown], outputs=audio_output)

if __name__ == "__main__":
    interface.launch()