BLAT Pre-trained Checkpoints and Data

Checkpoints of BLAT (Bootstrapping Language-Audio pre-training based on Tag-guided synthetic data) pre-trained models. The models are first pre-trained on synthetic AudioSet tag-guided audio-text data, then fine-tuned on human-annotated audio-text data (AudioCaps + Clotho + MACS).

Caption data generated by the AudioSet tag-guided captioning model is also provided.