Introduction
With the proliferation of music streaming services, the volume of user-generated and editorial playlists has reached billions. Organizing and discovering this content remains a significant challenge. While professional editors produce high-quality captions, the process is labor-intensive and difficult to scale. Automated captioning systems aim to bridge the "semantic gap"—the disconnect between low-level raw audio signals and high-level human concepts like "chill evening vibes" or "90s smash hits."
Previous automated approaches often relied on rigid templates or struggled with noisy, user-generated datasets. This paper introduces PLAYNTELL, a multi-modal model designed to generate natural language titles and descriptions for music playlists using limited "editorial-grade" data.
Methodology
The PLAYNTELL Architecture
PLAYNTELL utilizes an encoder-decoder architecture specifically engineered to integrate diverse musical and linguistic knowledge sources. The model's encoder consists of three specialized branches:
- Audio Branch: Employs a pre-trained CNN (VGGish) to extract acoustic features from 30-second track previews.
- Tag Branch: Leverages linguistic knowledge by extracting genre, mood, and era tags (e.g., "rock," "happy") from the Discogs database, providing a semantic bridge between audio and text.
- Artist Distribution Branch: Tracks artist dominance within a playlist to identify artist-centric themes (e.g., "100% The Beatles").
Data-Efficient Decoding
To address the scarcity of high-quality editorial data, the authors "warm-start" the decoder using pre-trained GPT-2 weights. This initialization provides the model with a foundational understanding of English syntax and grammar, allowing the training phase to focus on mapping musical context to linguistic descriptors. Additionally, an Artist Masking Strategy is used to handle the long tail of millions of artists by using popularity-based placeholders during training, which are replaced with real names during inference.
Experimental Setup and Datasets
The researchers curated a new dataset of editorial playlists from Deezer (5,467 playlists) and Spotify (1,104 playlists). Unlike user-generated datasets, these are curated by professionals, ensuring thematic consistency. These playlists were enriched with metadata from the Discogs database to provide high-level descriptors for the model.
Key Results
PLAYNTELL demonstrates a significant performance leap over existing baselines such as MUSCAPS and DOHRNN. Quantitative evaluations across multiple metrics show that the model achieves a 2x to 3x improvement in scores like BLEU@4 and CIDEr.
Beyond accuracy, PLAYNTELL exhibits superior diversity and novelty. It utilizes a significantly larger vocabulary (~2,500 words compared to ~800 for baselines) and generates a higher percentage of novel captions. In human evaluations conducted with professional editors, PLAYNTELL was rated significantly higher for "Content Match," confirming that its generated captions accurately reflect the musical themes of the playlists.
Discussion and Limitations
While PLAYNTELL sets a new state-of-the-art, it faces certain limitations. The current training data is heavily skewed toward Western music, leading to performance drops in regional genres like "Desi" music where tags are underrepresented. Furthermore, the model is currently limited to English caption generation.
Future research will focus on extending the model to support multi-lingual captioning and integrating these natural language descriptions into recommender systems to provide explainable music discovery for users.
Summary updated