Music Recommendation and Classification

Introduction

A project that utilizes Convolutional Neural Networks and feature selection techniques from audio data to create a music recommendation system and genre classification, providing a personalized and seamless listening experience.

Presentation

Explaination of our thinking

Data Preprocessing

Initial Steps

• Loading Audio Files

  • Utilize the librosa library to load audio files.

  • import librosay, sr = librosa.load('path\_to\_audio\_file.wav')

  • y is the audio time series, and sr is the sampling rate.

Transformation Process

• Fourier Transform(feature extraction)

  • Apply Fourier Transform to convert the time-domain audio signal into the frequency domain.

  • D = librosa.stft(y)

  • D is the short-time Fourier transform of the audio signal.

• Mel-frequency Cepstral Coefficients (MFCC) Extraction

  • Extract MFCC features which are critical for audio analysis and classification.

  • mfccs = librosa.feature.mfcc(y=y, sr=sr, n\_mfcc=13)

  • mfccs is a matrix containing the MFCC features.

Importance of Feature Extraction

• Feature Extraction

  • Essential for converting raw audio data into a format that is suitable.

  • Helps in capturing the relevant patterns and characteristics of the audio signal.

Model Architecture

---

• Convolutional Layers

  • Use multiple Conv2D layers to capture spatial features from the audio spectrograms.

  • Parameters include the number of filters, filter sizes, and strides.

  • from tensorflow.keras.layers import Conv2Dmodel.add(Conv2D(filters=32, kernel\_size=(3, 3), strides=(1, 1), activation='relu', input\_shape=input\_shape))

• Activation Functions

  • Primarily use ReLU (Rectified Linear Unit) for non-linearity.

  • from tensorflow.keras.layers import Activationmodel.add(Activation('relu'))

• Pooling Layers

  • Incorporate MaxPooling2D layers to downsample the feature maps.

  • from tensorflow.keras.layers import MaxPooling2Dmodel.add(MaxPooling2D(pool\_size=(2, 2)))

• Dropout Layers

  • Use Dropout layers to prevent overfitting by randomly setting a fraction of input units to 0.

  • from tensorflow.keras.layers import Dropoutmodel.add(Dropout(rate=0.5))

• Batch Normalization

  • Apply BatchNormalization to normalize the output of previous layers, accelerating training and improving performance.

  • from tensorflow.keras.layers import BatchNormalizationmodel.add(BatchNormalization())

• Flattening

  • Use Flatten to convert the 2D matrices into a 1D vector before passing to fully connected layers.

  • from tensorflow.keras.layers import Flattenmodel.add(Flatten())

Rationale and Design Philosophy

• Parameter Selection

  • Careful selection of parameters like the number of filters, kernel sizes, and dropout rates to balance performance and complexity.

  • Conv2D(filters=32, kernel\_size=(3, 3), strides=(1, 1))Dropout(rate=0.5)

  • Ensures that the model is both deep enough to capture complex patterns and regularized to prevent overfitting.

Made by:

Pranjal Vanjale

Siddhant Gupta

Vidhi Gupta

Yashovardhan Pandey

(Students of IIT Roorkee B.tech. First Year)

Mentored By: Shreshth Mehrotra

Ashwarya Rao Maratha