DuoTok: Source-Aware Dual-Track Tokenization for Multi-Track Music Language Modeling

1. 1Pretrain a semantic encoder to learn representations of audio data
2. 2Apply staged disentanglement to separate tokens into dual tracks
3. 3Evaluate the tokenizer on metrics such as reconstruction loss, perplexity, and cross-track correlation

AI engineers and researchers working on music language models can benefit from DuoTok's ability to preserve high-fidelity reconstruction and strong predictability, while also considering cross-track correspondence

💡 DuoTok's staged disentanglement approach allows for effective tokenization of multi-track music data while preserving important properties

🎵 Introducing DuoTok, a dual-track tokenizer for music language models that balances fidelity, predictability, and cross-track correspondence!