talkpl-ai/TalkPlayData-1

TalkPlay Dataset 🎵 💬

The TalkPlay dataset is a large-scale music conversation dataset created using the Million Playlist Dataset (MPD) as its foundation. The MPD, introduced in 2018, contains one million Spotify playlists and remains one of the largest publicly available playlist datasets. It provides comprehensive track metadata and playlist co-occurrence information. We leverage pretrained captioning, transcription, and MIR models to expand modalities, and transform playlist data into conversational data through LLMs.

Dataset Description 📊

The TalkPlay dataset was generated using GEMINI-1.5-FLASH-002, creating natural language conversations between users and an AI assistant. Key design principles include:

• Coherence in dialogue and music recommendations 🗣️
• Coverage of multiple modalities (audio, lyrics, metadata, semantic annotations) 🎼
  • Audio content crawled from Spotify API (1,717,148 items) 🎧
  • Lyrics extracted using Whisper-Large-V3 📝
  • Semantic annotations (genres, moods, instruments, vocal styles) from LP-MusicCaps 🏷️
• Realistic user simulation including track skipping/rejection 🔄
• Structured JSON format for training compatibility 💻

Dataset Statistics 📊

• 116k source playlists
• Average 7 tracks per playlist
• 552k multi-turn dialogues
• Expanded text-music associations (318k tokens vs original 24k)

Dataset	Training	Evaluation
Million Playlist Dataset
# of Playlists	999,000	1,000
# of Average Tracklist	63.73	59.41
# of Warm Start Tracks	1,714,772	37,114
# of Cold Start Tracks	-	2,357
# of Vocab	24,129	713
Sparsity	0.99996	0.99850
TalkPlay Dataset
# of Playlist	116,136	1,000
# of Dialogue	532,627	1,000
# of Tracks	406,768	11,022
# of Vocab	318,281	19,796
Avg. # of turns	7.03	13.31
Avg. query len.	17.62	15.35
Avg. response len.	17.91	14.92

Data Format 📋

Each example follows a triplet structure {USER QUERY, RECOMMENDED MUSIC, ASSISTANT RESPONSE} and contains:

• pid: Playlist ID (string) 🔑
• cid: Conversation ID (string) 🏷️
• conversations: List of messages, where each message has:
  • role: One of ["user", "assistant", "music"] 👥
  • content: Text content for user/assistant messages, Spotify track ID for music messages 💭

Spotify track metadata can be obtained from the official Million Playlist Dataset.(mapping with spotfiy unique ID)

License

The contents of this repository are made available under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International license.

Citation

@article{doh2025talkplay,
  title={TALKPLAY: Multimodal Music Recommendation with Large Language Models},
  author={Doh, Seungheon and Choi, Keunwoo and Nam, Juhan},
  journal={arXiv preprint arXiv:2502.13713},
  year={2025}
}