app.py

import gradio as gr
import os
import torch
from transformers import (
    AutoTokenizer,
    AutoModelForCausalLM,
    pipeline,
    AutoProcessor,
    MusicgenForConditionalGeneration,
)
from scipy.io.wavfile import write
from pydub import AudioSegment
from pydub.playback import play
import tempfile
from dotenv import load_dotenv
import spaces

# Load environment variables
load_dotenv()
hf_token = os.getenv("HF_TOKEN")

# ---------------------------------------------------------------------
# Script Generation Function
# ---------------------------------------------------------------------
@spaces.GPU(duration=300)
def generate_script(user_prompt: str, model_id: str, token: str, duration: int):
    try:
        tokenizer = AutoTokenizer.from_pretrained(model_id, use_auth_token=token)
        model = AutoModelForCausalLM.from_pretrained(
            model_id,
            use_auth_token=token,
            torch_dtype=torch.float16,
            device_map="auto",
            trust_remote_code=True,
        )
        llama_pipeline = pipeline("text-generation", model=model, tokenizer=tokenizer)

        system_prompt = (
            f"You are an expert radio imaging producer specializing in sound design and music. "
            f"Based on the user's concept and the selected duration of {duration} seconds, craft a concise, engaging promo script. "
            f"Ensure the script fits within the time limit and suggest a matching music style that complements the theme."
        )

        combined_prompt = f"{system_prompt}\nUser concept: {user_prompt}\nRefined script and music suggestion:"
        result = llama_pipeline(combined_prompt, max_new_tokens=200, do_sample=True, temperature=0.9)

        generated_text = result[0]["generated_text"].split("Refined script and music suggestion:")[-1].strip()
        script, music_suggestion = generated_text.split("Music Suggestion:")
        return script.strip(), music_suggestion.strip()
    except Exception as e:
        return f"Error generating script: {e}", None

# ---------------------------------------------------------------------
# Voice-Over Generation Function
# ---------------------------------------------------------------------
@spaces.GPU(duration=300)
def generate_voice(script: str, speaker: str):
    try:
        # Replace with your chosen TTS model
        tts_model = "coqui/XTTS-v2"
        processor = AutoProcessor.from_pretrained(tts_model)
        model = AutoModelForCausalLM.from_pretrained(tts_model)

        inputs = processor(script, return_tensors="pt")
        speech = model.generate(**inputs)

        output_path = f"{tempfile.gettempdir()}/generated_voice.wav"
        write(output_path, 22050, speech.cpu().numpy())
        return output_path
    except Exception as e:
        return f"Error generating voice-over: {e}"

# ---------------------------------------------------------------------
# Music Generation Function
# ---------------------------------------------------------------------
@spaces.GPU(duration=300)
def generate_music(prompt: str, audio_length: int):
    try:
        musicgen_model = MusicgenForConditionalGeneration.from_pretrained("facebook/musicgen-small")
        musicgen_processor = AutoProcessor.from_pretrained("facebook/musicgen-small")

        device = "cuda" if torch.cuda.is_available() else "cpu"
        musicgen_model.to(device)

        inputs = musicgen_processor(text=[prompt], padding=True, return_tensors="pt").to(device)
        outputs = musicgen_model.generate(**inputs, max_new_tokens=audio_length)

        audio_data = outputs[0, 0].cpu().numpy()
        normalized_audio = (audio_data / max(abs(audio_data)) * 32767).astype("int16")

        output_path = f"{tempfile.gettempdir()}/generated_music.wav"
        write(output_path, 44100, normalized_audio)

        return output_path
    except Exception as e:
        return f"Error generating music: {e}"

# ---------------------------------------------------------------------
# Audio Blending Function with Ducking
# ---------------------------------------------------------------------
def blend_audio(voice_path: str, music_path: str, ducking: bool):
    try:
        voice = AudioSegment.from_file(voice_path)
        music = AudioSegment.from_file(music_path)

        if ducking:
            music = music - 10  # Lower music volume for ducking

        combined = music.overlay(voice)
        output_path = f"{tempfile.gettempdir()}/final_promo.wav"
        combined.export(output_path, format="wav")

        return output_path
    except Exception as e:
        return f"Error blending audio: {e}"

# ---------------------------------------------------------------------
# Gradio Interface
# ---------------------------------------------------------------------
def process_all(user_prompt, llama_model_id, duration, audio_length, speaker, ducking):
    script, music_suggestion = generate_script(user_prompt, llama_model_id, hf_token, duration)
    if "Error" in script:
        return script, None

    voice_path = generate_voice(script, speaker)
    if "Error" in voice_path:
        return voice_path, None

    music_path = generate_music(music_suggestion, audio_length)
    if "Error" in music_path:
        return music_path, None

    final_audio = blend_audio(voice_path, music_path, ducking)
    return f"Script:\n{script}\n\nMusic Suggestion:\n{music_suggestion}", final_audio

with gr.Blocks() as demo:
    gr.Markdown("""
        # 🎧 AI Promo Studio with Script, Voice, Music, and Mixing 🚀  
        Generate fully mixed promos effortlessly with AI-driven tools for radio and media!
    """)

    with gr.Row():
        user_prompt = gr.Textbox(label="Promo Idea", placeholder="E.g., A 30-second promo for a morning show.")
        llama_model_id = gr.Textbox(label="Llama Model ID", value="meta-llama/Meta-Llama-3-8B-Instruct")
        duration = gr.Slider(label="Duration (seconds)", minimum=15, maximum=60, step=15, value=30)
        audio_length = gr.Slider(label="Music Length (tokens)", minimum=128, maximum=1024, step=64, value=512)
        speaker = gr.Textbox(label="Voice Style (optional)", placeholder="E.g., male, female, or neutral.")
        ducking = gr.Checkbox(label="Enable Ducking", value=True)

    generate_button = gr.Button("Generate Full Promo")
    script_output = gr.Textbox(label="Generated Script and Music Suggestion")
    audio_output = gr.Audio(label="Final Promo Audio", type="filepath")

    generate_button.click(
        fn=process_all,
        inputs=[user_prompt, llama_model_id, duration, audio_length, speaker, ducking],
        outputs=[script_output, audio_output],
    )

    gr.Markdown("""
        <hr>
        <p style="text-align: center; font-size: 0.9em;">
            Created with ❤️ by <a href="https://bilsimaging.com" target="_blank">bilsimaging.com</a>
        </p>
    """)

demo.launch(debug=True)