Song Identifier

A project that identifies songs and artists from lyric snippets using two distinct methods - simple NLP based approach and BM25(Best Match 25) approach.

1. Introduction

This project identifies songs and artists from lyric snippets using two distinct methods, tested on the Spotify dataset (57,650 songs) from Kaggle. It includes a Jupyter notebook (spotify_song_pred.ipynb) exploring both approaches, separated into standalone Python scripts (normal_nlp.py and bm25_approach.py) for ease of use.

Example queries like "sing us a song you're the piano man..." and "She's just my kind of girl..." are matched accurately, with BM25 achieving perfect confidence (1.00).

2. Approaches

Normal NLP (Hybrid TF-IDF + Cosine + Jaccard)

• Description: A hybrid approach combining TF-IDF vectorization with both Cosine and Jaccard similarity measures to optimize lyric matching.
• Performance: Correctly identifies songs (e.g., "Piano Man by Billy Joel" at 0.33 confidence), suitable for exact-match tasks.
• File: normal_nlp.py
• Technical Details:
  1. TF-IDF Vectorization:
     • vectorizer = TfidfVectorizer(ngram_range=(1,3)): Converts lyrics to TF-IDF vectors, capturing unigrams, bigrams, and trigrams (e.g., "piano," "piano man," "sing us song").
     • input_vector = vectorizer.transform([input_lyric]): Turns the query into a TF-IDF vector.
  2. Cosine Similarity:
     • cosine_scores = cosine_similarity(input_vector, tfidf_matrix).flatten(): Measures similarity between the query's TF-IDF vector and all song vectors, based on angle (cosine) between them.
     • Range: 0 to 1 (1 = identical).
  3. Jaccard Similarity:
     • jaccard_scores = np.array([jaccard_similarity(input_lyric, song_lyric) for song_lyric in cleaned_lyrics]): Calculates word overlap between the query and each song's cleaned text as a set intersection over union.
     • Range: 0 to 1 (1 = all words match).
  4. Combined Scores:
     • combined_scores = 0.7 * cosine_scores + 0.3 * jaccard_scores: Blends Cosine (70%) and Jaccard (30%) into a single score.
     • This combination leverages TF-IDF's weighted term importance through Cosine similarity, while Jaccard ensures exact word matches boost the score.

BM25 (via rank_bm25)

• Description: Uses the BM25Okapi algorithm from the rank_bm25 library, tuned with k1=1.5 and b=0.75, optimizing for retrieval by balancing term frequency and lyric length.
• Performance: Achieves perfect matches (e.g., "Piano Man by Billy Joel" and "She's My Kind Of Girl by ABBA" at 1.00 confidence), outperforming Normal NLP.
• File: bm25_approach.py

3. Setup

Cloning the Repository

git clone https://github.com/ArNAB-0053/Song-Identifier.git
cd Song-Identifier

Installing Dependencies

Manually install the required Python libraries, as no requirements.txt is provided (version compatibility may vary—use latest versions unless issues arise):

• For both approaches:

pip install pandas numpy nltk

• For Normal NLP:

pip install scikit-learn

• For BM25:

pip install rank_bm25

Dataset

• You can use the spotify_songs_dataset.csv file included in this repository.
• If you want the latest version (or if any updates occur), you can download it from Kaggle.
• Place the file in the project folder (alongside the scripts) before running the code.

4. Running the Code

Explore Both Approaches

• Open spotify_song_pred.ipynb in Jupyter Notebook to see the combined implementation and experimentation:

jupyter notebook spotify_song_pred.ipynb

Run Individual Scripts

• Normal NLP:

python normal_nlp.py

• BM25:

python bm25_approach.py

• Edit query, query1, or query2 in the scripts for custom snippets.