import os
import re
import torch
import tempfile
from scipy.io.wavfile import write
from pydub import AudioSegment
from dotenv import load_dotenv
import spaces
import gradio as gr
# Transformers & Models
from transformers import (
AutoTokenizer,
AutoModelForCausalLM,
pipeline,
AutoProcessor,
MusicgenForConditionalGeneration,
)
# Coqui TTS
from TTS.api import TTS
# ---------------------------------------------------------------------
# Load Environment Variables
# ---------------------------------------------------------------------
load_dotenv()
HF_TOKEN = os.getenv("HF_TOKEN")
# ---------------------------------------------------------------------
# Global Model Caches
# ---------------------------------------------------------------------
LLAMA_PIPELINES = {}
MUSICGEN_MODELS = {}
TTS_MODELS = {}
# ---------------------------------------------------------------------
# Utility Function: Clean Text
# ---------------------------------------------------------------------
def clean_text(text: str) -> str:
"""
Removes undesired characters (e.g., asterisks) that might not be recognized by the model's vocabulary.
"""
# Remove all asterisks. You can add more cleaning steps here as needed.
return re.sub(r'\*', '', text)
# ---------------------------------------------------------------------
# Helper Functions
# ---------------------------------------------------------------------
def get_llama_pipeline(model_id: str, token: str):
"""
Returns a cached LLaMA pipeline if available; otherwise, loads it.
"""
if model_id in LLAMA_PIPELINES:
return LLAMA_PIPELINES[model_id]
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,
)
text_pipeline = pipeline("text-generation", model=model, tokenizer=tokenizer)
LLAMA_PIPELINES[model_id] = text_pipeline
return text_pipeline
def get_musicgen_model(model_key: str = "facebook/musicgen-large"):
"""
Returns a cached MusicGen model if available; otherwise, loads it.
Uses the 'large' variant for higher quality outputs.
"""
if model_key in MUSICGEN_MODELS:
return MUSICGEN_MODELS[model_key]
model = MusicgenForConditionalGeneration.from_pretrained(model_key)
processor = AutoProcessor.from_pretrained(model_key)
device = "cuda" if torch.cuda.is_available() else "cpu"
model.to(device)
MUSICGEN_MODELS[model_key] = (model, processor)
return model, processor
def get_tts_model(model_name: str = "tts_models/en/ljspeech/tacotron2-DDC"):
"""
Returns a cached TTS model if available; otherwise, loads it.
"""
if model_name in TTS_MODELS:
return TTS_MODELS[model_name]
tts_model = TTS(model_name)
TTS_MODELS[model_name] = tts_model
return tts_model
# ---------------------------------------------------------------------
# Script Generation Function
# ---------------------------------------------------------------------
@spaces.GPU(duration=100)
def generate_script(user_prompt: str, model_id: str, token: str, duration: int):
"""
Generates a script, sound design suggestions, and music ideas from a user prompt.
Returns a tuple of strings: (voice_script, sound_design, music_suggestions).
"""
try:
text_pipeline = get_llama_pipeline(model_id, token)
system_prompt = (
"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, produce the following: "
"1. A concise voice-over script. Prefix this section with 'Voice-Over Script:'.\n"
"2. Suggestions for sound design. Prefix this section with 'Sound Design Suggestions:'.\n"
"3. Music styles or track recommendations. Prefix this section with 'Music Suggestions:'."
)
combined_prompt = f"{system_prompt}\nUser concept: {user_prompt}\nOutput:"
with torch.inference_mode():
result = text_pipeline(
combined_prompt,
max_new_tokens=300,
do_sample=True,
temperature=0.8
)
generated_text = result[0]["generated_text"]
if "Output:" in generated_text:
generated_text = generated_text.split("Output:")[-1].strip()
# Default placeholders
voice_script = "No voice-over script found."
sound_design = "No sound design suggestions found."
music_suggestions = "No music suggestions found."
# Voice-Over Script
if "Voice-Over Script:" in generated_text:
parts = generated_text.split("Voice-Over Script:")
voice_script_part = parts[1]
if "Sound Design Suggestions:" in voice_script_part:
voice_script = voice_script_part.split("Sound Design Suggestions:")[0].strip()
else:
voice_script = voice_script_part.strip()
# Sound Design
if "Sound Design Suggestions:" in generated_text:
parts = generated_text.split("Sound Design Suggestions:")
sound_design_part = parts[1]
if "Music Suggestions:" in sound_design_part:
sound_design = sound_design_part.split("Music Suggestions:")[0].strip()
else:
sound_design = sound_design_part.strip()
# Music Suggestions
if "Music Suggestions:" in generated_text:
parts = generated_text.split("Music Suggestions:")
music_suggestions = parts[1].strip()
return voice_script, sound_design, music_suggestions
except Exception as e:
return f"Error generating script: {e}", "", ""
# ---------------------------------------------------------------------
# Voice-Over Generation Function
# ---------------------------------------------------------------------
@spaces.GPU(duration=100)
def generate_voice(script: str, tts_model_name: str = "tts_models/en/ljspeech/tacotron2-DDC"):
"""
Generates a voice-over from the provided script using the Coqui TTS model.
Returns the file path to the generated .wav file.
"""
try:
if not script.strip():
return "Error: No script provided."
# Clean the script to remove special characters (e.g., asterisks) that may produce warnings
cleaned_script = clean_text(script)
tts_model = get_tts_model(tts_model_name)
# Generate and save voice
output_path = os.path.join(tempfile.gettempdir(), "voice_over.wav")
tts_model.tts_to_file(text=cleaned_script, file_path=output_path)
return output_path
except Exception as e:
return f"Error generating voice: {e}"
# ---------------------------------------------------------------------
# Music Generation Function
# ---------------------------------------------------------------------
@spaces.GPU(duration=200)
def generate_music(prompt: str, audio_length: int):
"""
Generates music from the 'facebook/musicgen-large' model based on the prompt.
Returns the file path to the generated .wav file.
"""
try:
if not prompt.strip():
return "Error: No music suggestion provided."
model_key = "facebook/musicgen-large"
musicgen_model, musicgen_processor = get_musicgen_model(model_key)
device = "cuda" if torch.cuda.is_available() else "cpu"
inputs = musicgen_processor(text=[prompt], padding=True, return_tensors="pt").to(device)
with torch.inference_mode():
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 = os.path.join(tempfile.gettempdir(), "musicgen_large_generated_music.wav")
write(output_path, 44100, normalized_audio)
return output_path
except Exception as e:
return f"Error generating music: {e}"
# ---------------------------------------------------------------------
# Audio Blending with Duration Sync & Ducking
# ---------------------------------------------------------------------
@spaces.GPU(duration=100)
def blend_audio(voice_path: str, music_path: str, ducking: bool, duck_level: int = 10):
"""
Blends two audio files (voice and music).
1. If music < voice, loops the music until it meets/exceeds the voice duration.
2. If music > voice, trims music to the voice duration.
3. If ducking=True, the music is attenuated by 'duck_level' dB while the voice is playing.
Returns the file path to the blended .wav file.
"""
try:
if not os.path.isfile(voice_path) or not os.path.isfile(music_path):
return "Error: Missing audio files for blending."
voice = AudioSegment.from_wav(voice_path)
music = AudioSegment.from_wav(music_path)
voice_len = len(voice) # in milliseconds
music_len = len(music) # in milliseconds
# Loop music if it's shorter than the voice
if music_len < voice_len:
looped_music = AudioSegment.empty()
while len(looped_music) < voice_len:
looped_music += music
music = looped_music
# Trim music if it's longer than the voice
if len(music) > voice_len:
music = music[:voice_len]
if ducking:
ducked_music = music - duck_level
final_audio = ducked_music.overlay(voice)
else:
final_audio = music.overlay(voice)
output_path = os.path.join(tempfile.gettempdir(), "blended_output.wav")
final_audio.export(output_path, format="wav")
return output_path
except Exception as e:
return f"Error blending audio: {e}"
# ---------------------------------------------------------------------
# Gradio Interface with Enhanced UI
# ---------------------------------------------------------------------
with gr.Blocks(css="""
/* Global Styles */
body {
background: linear-gradient(135deg, #1d1f21, #3a3d41);
color: #f0f0f0;
font-family: 'Segoe UI', Tahoma, Geneva, Verdana, sans-serif;
}
.header {
text-align: center;
padding: 2rem 1rem;
background: linear-gradient(90deg, #6a11cb, #2575fc);
border-radius: 0 0 20px 20px;
margin-bottom: 2rem;
}
.header h1 {
margin: 0;
font-size: 2.5rem;
}
.header p {
font-size: 1.2rem;
}
.gradio-container {
background: #2e2e2e;
border-radius: 10px;
padding: 1rem;
}
.tab-title {
font-size: 1.1rem;
font-weight: bold;
}
.footer {
text-align: center;
font-size: 0.9em;
margin-top: 2rem;
padding: 1rem;
color: #cccccc;
}
""") as demo:
# Custom Header
with gr.Row(elem_classes="header"):
gr.Markdown("""
🎧 AI Promo Studio
Your all-in-one AI solution for crafting engaging audio promos.
""")
gr.Markdown("""
Welcome to **AI Promo Studio**! This platform leverages state-of-the-art AI models to help you generate:
- **Script**: Generate a compelling voice-over script with LLaMA.
- **Voice Synthesis**: Create natural-sounding voice-overs using Coqui TTS.
- **Music Production**: Produce custom music tracks with MusicGen.
- **Audio Blending**: Seamlessly blend voice and music with options for ducking.
""")
with gr.Tabs():
# Step 1: Generate Script
with gr.Tab("📝 Script Generation"):
with gr.Row():
user_prompt = gr.Textbox(
label="Promo Idea",
placeholder="E.g., A 30-second promo for a morning show...",
lines=2
)
with gr.Row():
llama_model_id = gr.Textbox(
label="LLaMA Model ID",
value="meta-llama/Meta-Llama-3-8B-Instruct",
placeholder="Enter a valid Hugging Face model ID"
)
duration = gr.Slider(
label="Desired Promo Duration (seconds)",
minimum=15,
maximum=60,
step=15,
value=30
)
generate_script_button = gr.Button("Generate Script", variant="primary")
script_output = gr.Textbox(label="Generated Voice-Over Script", lines=5, interactive=False)
sound_design_output = gr.Textbox(label="Sound Design Suggestions", lines=3, interactive=False)
music_suggestion_output = gr.Textbox(label="Music Suggestions", lines=3, interactive=False)
generate_script_button.click(
fn=lambda user_prompt, model_id, dur: generate_script(user_prompt, model_id, HF_TOKEN, dur),
inputs=[user_prompt, llama_model_id, duration],
outputs=[script_output, sound_design_output, music_suggestion_output],
)
# Step 2: Generate Voice
with gr.Tab("🎤 Voice Synthesis"):
gr.Markdown("Generate a natural-sounding voice-over using Coqui TTS.")
selected_tts_model = gr.Dropdown(
label="TTS Model",
choices=[
"tts_models/en/ljspeech/tacotron2-DDC",
"tts_models/en/ljspeech/vits",
"tts_models/en/sam/tacotron-DDC",
],
value="tts_models/en/ljspeech/tacotron2-DDC",
multiselect=False
)
generate_voice_button = gr.Button("Generate Voice-Over", variant="primary")
voice_audio_output = gr.Audio(label="Voice-Over (WAV)", type="filepath")
generate_voice_button.click(
fn=lambda script, tts_model: generate_voice(script, tts_model),
inputs=[script_output, selected_tts_model],
outputs=voice_audio_output,
)
# Step 3: Generate Music
with gr.Tab("🎶 Music Production"):
gr.Markdown("Generate a custom music track using the **MusicGen Large** model.")
audio_length = gr.Slider(
label="Music Length (tokens)",
minimum=128,
maximum=1024,
step=64,
value=512,
info="Increase tokens for longer audio (inference time may vary)."
)
generate_music_button = gr.Button("Generate Music", variant="primary")
music_output = gr.Audio(label="Generated Music (WAV)", type="filepath")
generate_music_button.click(
fn=lambda music_suggestion, length: generate_music(music_suggestion, length),
inputs=[music_suggestion_output, audio_length],
outputs=[music_output],
)
# Step 4: Blend Audio
with gr.Tab("🎚️ Audio Blending"):
gr.Markdown("Blend your voice-over and music track. Music will be looped/truncated to match the voice duration. Enable ducking to lower the music during voice segments.")
ducking_checkbox = gr.Checkbox(label="Enable Ducking?", value=True)
duck_level_slider = gr.Slider(
label="Ducking Level (dB attenuation)",
minimum=0,
maximum=20,
step=1,
value=10
)
blend_button = gr.Button("Blend Voice + Music", variant="primary")
blended_output = gr.Audio(label="Final Blended Output (WAV)", type="filepath")
blend_button.click(
fn=blend_audio,
inputs=[voice_audio_output, music_output, ducking_checkbox, duck_level_slider],
outputs=blended_output
)
# Footer
gr.Markdown("""
""")
# Visitor Badge
gr.HTML("""
""")
demo.launch(debug=True)