app.py

import gradio as gr
from transformers import pipeline
import torch
from transformers import AutoProcessor, MusicgenForConditionalGeneration
import numpy as np # Import numpy

# Load emotion classifier
emotion_classifier = pipeline("text-classification", model="j-hartmann/emotion-english-distilroberta-base", return_all_scores=True)

# Load music generator (small for CPU)
music_model = MusicgenForConditionalGeneration.from_pretrained("facebook/musicgen-small")
processor = AutoProcessor.from_pretrained("facebook/musicgen-small")

# Map emotion to style/genre prompts
EMOTION_TO_MUSIC = {
    "joy": "happy upbeat piano melody",
    "anger": "intense aggressive drums",
    "sadness": "slow emotional violin",
    "fear": "dark ambient synth",
    "love": "soft romantic acoustic guitar",
    "surprise": "quirky playful tune",
    "neutral": "chill background lofi beat"
}

# Main generation function
def generate_music(user_input):
    # Step 1: Detect emotion
    emotion_scores = emotion_classifier(user_input)[0]
    top_emotion = max(emotion_scores, key=lambda x: x["score"])["label"]

    # Step 2: Generate prompt
    music_prompt = EMOTION_TO_MUSIC.get(top_emotion.lower(), "ambient melody")

    # Step 3: Generate music
    inputs = processor(text=[music_prompt], return_tensors="pt")
    audio_values = music_model.generate(**inputs, max_new_tokens=1024)

    # Convert audio tensor to numpy array
    audio_array = audio_values[0].cpu().numpy()

    # --- FIX START ---
    # Normalize the audio array to be within the range of a 16-bit PCM WAV file
    # The default sampling rate for musicgen-small is 16000 Hz, and Gradio expects
    # values to be scaled for 16-bit integers if not float.
    # We'll normalize to -1 to 1 for float and let Gradio handle the 16-bit conversion.
    # However, to be extra safe, ensure max amplitude is 1.0.
    audio_array = audio_array / np.max(np.abs(audio_array))
    # --- FIX END ---


    # Return result
    # The Musicgen model outputs audio at a sampling rate of 16kHz
    sampling_rate = 16000
    return f"Top Emotion: {top_emotion}", (sampling_rate, audio_array)

# Gradio UI
with gr.Blocks() as demo:
    gr.Markdown("# Emotion-to-Music AI")
    gr.Markdown("Describe how you feel and get a unique music track matching your mood!")

    with gr.Row():
        text_input = gr.Textbox(label="How are you feeling?")
        generate_btn = gr.Button("Generate Music")

    with gr.Row():
        emotion_output = gr.Textbox(label="Detected Emotion")
        audio_output = gr.Audio(label="Generated Music", type="numpy") # type="numpy" is correct here

    generate_btn.click(fn=generate_music, inputs=text_input, outputs=[emotion_output, audio_output])

demo.launch()