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 dotenv import load_dotenv
import tempfile
import spaces
from TTS.api import TTS

# -------------------------------
# Configuration
# -------------------------------
load_dotenv()
HF_TOKEN = os.getenv("HF_TOKEN")

MODEL_CONFIG = {
    "llama_models": {
        "Meta-Llama-3-8B": "meta-llama/Meta-Llama-3-8B-Instruct",
        "Mistral-7B": "mistralai/Mistral-7B-Instruct-v0.2",
        "Phi-3-mini": "microsoft/Phi-3-mini-4k-instruct"
    },
    "tts_models": {
        "Standard English": "tts_models/en/ljspeech/tacotron2-DDC",
        "High Quality": "tts_models/en/ljspeech/vits",
        "Fast Inference": "tts_models/en/sam/tacotron-DDC"
    }
}

# -------------------------------
# Model Manager
# -------------------------------
class ModelManager:
    def __init__(self):
        self.llama_pipelines = {}
        self.musicgen_models = {}
        self.tts_models = {}

    def get_llama_pipeline(self, model_id, token):
        if model_id not in self.llama_pipelines:
            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",
                attn_implementation="flash_attention_2"
            )
            self.llama_pipelines[model_id] = pipeline(
                "text-generation",
                model=model,
                tokenizer=tokenizer,
                device_map="auto"
            )
        return self.llama_pipelines[model_id]

    def get_musicgen_model(self, model_key="facebook/musicgen-large"):
        if model_key not in self.musicgen_models:
            model = MusicgenForConditionalGeneration.from_pretrained(model_key)
            processor = AutoProcessor.from_pretrained(model_key)
            device = "cuda" if torch.cuda.is_available() else "cpu"
            model.to(device)
            self.musicgen_models[model_key] = (model, processor)
        return self.musicgen_models[model_key]

    def get_tts_model(self, model_name):
        if model_name not in self.tts_models:
            self.tts_models[model_name] = TTS(model_name)
        return self.tts_models[model_name]

model_manager = ModelManager()

# -------------------------------
# Core Functions
# -------------------------------
@spaces.GPU(duration=120)
def generate_script(user_prompt, model_id, duration, temperature=0.7, max_tokens=512):
    try:
        text_pipeline = model_manager.get_llama_pipeline(model_id, HF_TOKEN)
        
        system_prompt = f"""You are an expert radio imaging producer. Create content for a {duration}-second promo:
1. Voice Script: [Clear narration]
2. Sound Design: [3-5 effects]
3. Music: [Genre/tempo/mood]

Respond in this exact format:"""

        prompt = f"{system_prompt}\nConcept: {user_prompt}\nVoice Script:"
        
        response = text_pipeline(
            prompt,
            max_new_tokens=max_tokens,
            temperature=temperature,
            do_sample=True,
            top_p=0.95,
            eos_token_id=text_pipeline.tokenizer.eos_token_id
        )

        return parse_generated_content(response[0]["generated_text"])
    except Exception as e:
        return f"Error: {str(e)}", "", ""

def parse_generated_content(text):
    sections = {
        "Voice Script": "",
        "Sound Design": "",
        "Music": ""
    }
    current_section = None
    
    for line in text.split('\n'):
        line = line.strip()
        if "Voice Script:" in line:
            current_section = "Voice Script"
            line = line.replace("Voice Script:", "").strip()
        elif "Sound Design:" in line:
            current_section = "Sound Design"
            line = line.replace("Sound Design:", "").strip()
        elif "Music:" in line:
            current_section = "Music"
            line = line.replace("Music:", "").strip()
        
        if current_section and line:
            sections[current_section] += line + "\n"
    
    return sections["Voice Script"].strip(), sections["Sound Design"].strip(), sections["Music"].strip()

@spaces.GPU(duration=100)
def generate_voice(script, tts_model, speed=1.0):
    try:
        if not script.strip():
            return "Error: Empty script"
            
        tts = model_manager.get_tts_model(tts_model)
        output_path = os.path.join(tempfile.gettempdir(), "voice.wav")
        
        tts.tts_to_file(
            text=script,
            file_path=output_path,
            speed=speed
        )
        return output_path
    except Exception as e:
        return f"Error: {str(e)}"

@spaces.GPU(duration=150)
def generate_music(prompt, duration_sec=30, temperature=1.0, guidance_scale=3.0):
    try:
        model, processor = model_manager.get_musicgen_model()
        device = "cuda" if torch.cuda.is_available() else "cpu"
        
        inputs = processor(
            text=[prompt],
            padding=True,
            return_tensors="pt",
        ).to(device)

        audio_values = model.generate(
            **inputs,
            max_new_tokens=int(duration_sec * 50),
            temperature=temperature,
            guidance_scale=guidance_scale,
            do_sample=True
        )

        output_path = os.path.join(tempfile.gettempdir(), "music.wav")
        write(output_path, 44100, audio_values[0, 0].cpu().numpy())
        return output_path
    except Exception as e:
        return f"Error: {str(e)}"

def blend_audio(voice_path, music_path, ducking=True, duck_level=10, crossfade=500):
    try:
        voice = AudioSegment.from_wav(voice_path)
        music = AudioSegment.from_wav(music_path)
        
        # Align durations with crossfade
        if len(music) < len(voice):
            loops = (len(voice) // len(music)) + 1
            music = music * loops
        
        music = music[:len(voice)].fade_out(crossfade)
        voice = voice.fade_in(crossfade)
        
        # Apply ducking
        if ducking:
            ducked_music = music - duck_level
            mixed = ducked_music.overlay(voice)
        else:
            mixed = music.overlay(voice)
            
        output_path = os.path.join(tempfile.gettempdir(), "final_mix.wav")
        mixed.export(output_path, format="wav")
        return output_path
    except Exception as e:
        return f"Error: {str(e)}"

# -------------------------------
# Gradio Interface
# -------------------------------
theme = gr.themes.Soft(
    primary_hue="blue",
    secondary_hue="teal",
).set(
    body_text_color_dark='#FFFFFF',
    background_fill_primary_dark='#1F1F1F'
)

with gr.Blocks(theme=theme, title="AI Radio Studio Pro") as demo:
    gr.Markdown("""
    # 🎧 AI Radio Studio Pro
    *Professional Audio Production in 4 Steps*
    """)

    with gr.Tabs():
        # Step 1: Concept Development
        with gr.Tab("1️⃣ Concept"):
            with gr.Row():
                with gr.Column(scale=2):
                    concept_input = gr.Textbox(
                        label="Your Idea",
                        placeholder="e.g., A 30-second morning show intro with energetic music...",
                        lines=3
                    )
                    with gr.Accordion("Advanced Settings", open=False):
                        model_selector = gr.Dropdown(
                            choices=list(MODEL_CONFIG["llama_models"].values()),
                            label="AI Model",
                            value=MODEL_CONFIG["llama_models"]["Meta-Llama-3-8B"]
                        )
                        duration_slider = gr.Slider(15, 120, 30, step=15, label="Duration (seconds)")
                        temp_slider = gr.Slider(0.1, 1.5, 0.7, step=0.1, label="Creativity")
                    generate_btn = gr.Button("Generate Script", variant="primary")
                
                with gr.Column(scale=1):
                    script_output = gr.Textbox(label="Voice Script", interactive=True)
                    sound_output = gr.Textbox(label="Sound Design", interactive=True)
                    music_output = gr.Textbox(label="Music Style", interactive=True)

        # Step 2: Voice Production
        with gr.Tab("2️⃣ Voice"):
            with gr.Row():
                with gr.Column():
                    tts_selector = gr.Dropdown(
                        choices=list(MODEL_CONFIG["tts_models"].values()),
                        label="Voice Model",
                        value="tts_models/en/ljspeech/tacotron2-DDC"
                    )
                    speed_slider = gr.Slider(0.5, 2.0, 1.0, step=0.1, label="Speaking Rate")
                    voice_btn = gr.Button("Generate Voiceover", variant="primary")
                with gr.Column():
                    voice_preview = gr.Audio(label="Preview", type="filepath")

        # Step 3: Music Production
        with gr.Tab("3️⃣ Music"):
            with gr.Row():
                with gr.Column():
                    music_duration = gr.Slider(10, 120, 30, label="Duration (seconds)")
                    music_temp = gr.Slider(0.1, 2.0, 1.0, label="Creativity")
                    guidance_scale = gr.Slider(1.0, 5.0, 3.0, label="Focus")
                    music_btn = gr.Button("Generate Music", variant="primary")
                with gr.Column():
                    music_preview = gr.Audio(label="Preview", type="filepath")

        # Step 4: Final Mix
        with gr.Tab("4️⃣ Mix"):
            with gr.Row():
                with gr.Column():
                    ducking_toggle = gr.Checkbox(True, label="Enable Voice Ducking")
                    duck_level = gr.Slider(0, 30, 12, label="Ducking Strength (dB)")
                    crossfade_time = gr.Slider(0, 2000, 500, label="Crossfade (ms)")
                    mix_btn = gr.Button("Create Final Mix", variant="primary")
                with gr.Column():
                    final_mix = gr.Audio(label="Master Output", type="filepath")

    # Examples & Footer
    with gr.Accordion("💡 Example Prompts", open=False):
        gr.Examples(
            examples=[
                ["A 45-second tech podcast intro with futuristic synth effects"],
                ["A 15-second coffee shop radio ad with morning acoustic vibes"],
                ["A 60-second documentary trailer with epic orchestral music"]
            ],
            inputs=concept_input
        )

    gr.Markdown("""
    <div style="text-align: center; margin-top: 30px; padding-top: 20px; border-top: 1px solid #444;">
        <p style="font-size: 0.9em; color: #888;">
            Created with ❤️ by <a href="https://bilsimaging.com" target="_blank" style="color: #66b3ff;">bilsimaging.com</a>
        </p>
        <a href="https://visitorbadge.io/status?path=https%3A%2F%2Fhuggingface.co%2Fspaces%2FBils%2Fradiogold">
            <img src="https://api.visitorbadge.io/api/visitors?path=https%3A%2F%2Fhuggingface.co%2Fspaces%2FBils%2Fradiogold&countColor=%23263759"/>
        </a>
    </div>
    """)

    # Event Handling
    generate_btn.click(
        generate_script,
        inputs=[concept_input, model_selector, duration_slider, temp_slider],
        outputs=[script_output, sound_output, music_output]
    )
    
    voice_btn.click(
        generate_voice,
        inputs=[script_output, tts_selector, speed_slider],
        outputs=voice_preview
    )
    
    music_btn.click(
        generate_music,
        inputs=[music_output, music_duration, music_temp, guidance_scale],
        outputs=music_preview
    )
    
    mix_btn.click(
        blend_audio,
        inputs=[voice_preview, music_preview, ducking_toggle, duck_level, crossfade_time],
        outputs=final_mix
    )

if __name__ == "__main__":
    demo.launch(server_name="0.0.0.0", server_port=7860)