bosonai/AudioTokenBench

AudioTokenBench是一个评估数据集，用于HiggsTokenizer。它包含3150个24khz的音频样本，分为四个子集：语音（Speech）、音乐（Music）、声音事件（Sound Event）和音响爱好者（Audiophile）。语音子集包含从DAPS随机抽取的1000个10秒音频片段；音乐子集包含从MUSDB随机抽取的1000个10秒音频片段；声音事件子集包含从AudioSet随机抽取的1000个10秒音频片段；音响爱好者子集包含从11张高保真测试光盘精选的150个30秒音频片段，这些片段包含音乐和声音事件，用于高品质音频评估。AudioTokenBench is an evaluation dataset for HiggsTokenizer. It contains 3150 24khz audio samples across four subsets: Speech, Music, Sound Event, and Audiophile. The Speech subset includes 1,000 10-second audio clips randomly sampled from DAPS; the Music subset includes 1,000 10-second audio clips randomly sampled from MUSDB; the Sound Event subset includes 1,000 10-second audio clips randomly sampled from AudioSet; the Audiophile subset contains 150 30-second audio clips curated from eleven high-fidelity test discs, featuring both music and sound events selected for high-quality audio evaluation.