import gradio as gr import numpy as np import librosa import tensorflow as tf from pathlib import Path # Genres based on your notebook’s mapping CLASS_NAMES = [ "blues","classical","country","disco","hiphop", "jazz","metal","pop","reggae","rock" ] MODEL_PATH = Path("model_4_complete.h5") model = tf.keras.models.load_model(MODEL_PATH) def extract_features(y, sr): if y.ndim > 1: y = librosa.to_mono(y) target_len = sr * 30 if len(y) < target_len: y = np.pad(y, (0, target_len - len(y))) else: y = y[:target_len] # Recreate the same features as in training CSV chroma_stft = librosa.feature.chroma_stft(y=y, sr=sr).mean() rms = librosa.feature.rms(y=y).mean() spec_cent = librosa.feature.spectral_centroid(y=y, sr=sr).mean() spec_bw = librosa.feature.spectral_bandwidth(y=y, sr=sr).mean() rolloff = librosa.feature.spectral_rolloff(y=y, sr=sr).mean() zcr = librosa.feature.zero_crossing_rate(y).mean() harmony = np.mean(np.abs(librosa.effects.harmonic(y))) flatness = librosa.feature.spectral_flatness(y=y).mean() tempi = librosa.beat.tempo(y=y, sr=sr, aggregate=None) tempo = float(np.median(tempi)) if tempi is not None and len(tempi) else 0.0 mfccs = librosa.feature.mfcc(y=y, sr=sr, n_mfcc=20).mean(axis=1) feats = np.hstack([ chroma_stft, rms, spec_cent, spec_bw, rolloff, zcr, harmony, flatness, tempo, mfccs ]).astype(np.float32) return feats def predict(file_path): y, sr = librosa.load(file_path, sr=22050, mono=True) feats = extract_features(y, sr)[None, :] probs = model.predict(feats, verbose=0)[0] idxs = probs.argsort()[-3:][::-1] return {CLASS_NAMES[i]: float(probs[i]) for i in idxs} demo = gr.Interface( fn=predict, inputs=gr.Audio(type="filepath", label="Upload a 5–30s clip (wav/mp3)"), outputs=gr.Label(num_top_classes=3, label="Top-3 Genres"), title="Music Genre Classification", description="Upload an audio clip and get predicted genres." ) if __name__ == "__main__": demo.launch()