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
import csv
from collections import defaultdict

# Model dictionary
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
    }
}

# Cache
MODEL_CACHE = {}
SPEAKER_DROPDOWN_MAP = {}  # Maps label -> ID

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, metadata_path="metadata.csv"):
    global SPEAKER_DROPDOWN_MAP
    info = VOICE_MODELS[model_key]
    if not info["multi_speaker"]:
        return []

    tts = load_tts_model(model_key)
    available_speakers = set(getattr(tts, "speakers", []))

    speaker_audio_map = defaultdict(list)
    with open(metadata_path, newline='') as csvfile:
        reader = csv.reader(csvfile)
        next(reader)
        for row in reader:
            if len(row) >= 2:
                audio_id, speaker_id = row[0], row[1]
                if speaker_id in available_speakers:
                    speaker_audio_map[speaker_id].append(audio_id)

    SPEAKER_DROPDOWN_MAP.clear()
    dropdown_choices = []
    for speaker_id, audio_ids in speaker_audio_map.items():
        label = f"{speaker_id} ({len(audio_ids)} samples)"
        SPEAKER_DROPDOWN_MAP[label] = speaker_id
        dropdown_choices.append(label)

    return dropdown_choices

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

    # Extract text
    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"] and speaker_label:
        speaker_id = SPEAKER_DROPDOWN_MAP.get(speaker_label)
        if not speaker_id:
            raise ValueError("Speaker ID not found.")
        kwargs["speaker"] = speaker_id

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

def show_load_button(voice_selection):
    is_multi = VOICE_MODELS[voice_selection]["multi_speaker"]
    return gr.update(visible=is_multi)

def load_speakers_ui(voice_selection):
    speakers = extract_speakers(voice_selection)
    return gr.update(choices=speakers, visible=True, value=speakers[0] if speakers else None)

with gr.Blocks() as interface:
    gr.Markdown("# 🗣️ DOCX to Realistic Voiceover")

    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_button = gr.Button("🔄 Load Speakers", visible=False)
        speaker_dropdown = gr.Dropdown(label="Speaker", visible=False)

    generate_button = gr.Button("🎙️ Generate Speech")
    audio_output = gr.Audio(label="🔊 Output WAV", type="filepath")

    # Event bindings
    voice_dropdown.change(fn=show_load_button, inputs=voice_dropdown, outputs=load_button)
    load_button.click(fn=load_speakers_ui, inputs=voice_dropdown, outputs=speaker_dropdown)

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

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