Clap Htsat Base

Clap Htsat Base is a model package available in ONNX format, designed for music audio and text embedding tasks. The bundle includes pre-converted ONNX Runtime artifacts based on the LAION-CLAP music checkpoint, specifically using the HTSAT-base architecture for audio and a RoBERTa-base architecture for text. The audio component accepts mono waveform inputs of 10 seconds at 48 kHz, processes them through the HTSAT-base audio tower, and outputs L2-normalized embeddings of size 512. The mel spectrogram is computed within the model graph. The text component utilizes the RoBERTa-base text tower, taking tokenized text inputs and producing similarly L2-normalized embeddings of size 512. The package also contains the exact RoBERTa tokenizer used by the checkpoint, ensuring consistency in text processing, as well as a manifest file with SHA-256 checksums and byte counts for each artifact. This model bundle is licensed under CC0-1.0, reflecting its open-source status and allowing for unrestricted use. It is intended as a mechanical ONNX conversion of the original upstream checkpoint and is packaged for optional use with a semantic-analysis provider named muzaiten-features-clap. The evidence does not specify particular user roles or application domains beyond its focus on music audio and text embedding, nor does it detail integration options or further deployment instructions. The model is hosted on Hugging Face and is available for download and use in compatible ONNX Runtime environments.