HookTheory

Sheet Sage是在一个新的数据集上训练的，该数据集包含50小时的对齐旋律和声注释，这些注释来自Hooktheory的TheoryTab DB，我们在CC BY-NC-SA 3.0许可证下与该系统一起发布。可以在此处以简单，MIR友好的JSON格式 (20mb) 下载数据集 (不包括音频)。单击此处查看独立的IPython笔记本，演示如何浏览数据集。 数据集是一个简单的JSON对象，其中每个注释都由其HookTheory ID键入。我们以按艺术家姓名分层的8:1:1比率将数据 (请参见拆分字段) 预拆分为训练，验证和测试子集。标签字段包含各种高级标签; 对于训练melody转录模型，我们建议向下过滤到包含audio_valified和MELODY标签的注释，并过滤掉包含TEMPO_CHANGES标签的注释。在 “对齐” 字段中，我们包括HookTheory的原始用户指定对齐方式和我们的改进对齐方式 (有关详细信息，请参阅我们的论文); 您的系统在培训过程中可能会使用其中之一 (或两者都不使用!)。Sheet Sage was trained on a novel dataset containing 50 hours of aligned melody and harmony annotations sourced from Hooktheory’s TheoryTab DB, which we release alongside the system under the CC BY-NC-SA 3.0 license. The dataset (excluding audio) can be downloaded here in a simple, MIR-friendly JSON format (20MB). Click here to view a standalone IPython notebook that demonstrates how to browse the dataset. The dataset is a simple JSON object, where each annotation is keyed by its HookTheory ID. We pre-split the data into training, validation, and test subsets with a stratified 8:1:1 ratio based on artist name (see the split field). The tags field contains various high-level labels; for training melody transcription models, we recommend filtering down to annotations that include the audio_valified and MELODY tags, and filtering out annotations that include the TEMPO_CHANGES tag. In the "alignment" field, we include both Hooktheory’s original user-specified alignments and our improved alignments (for details, please refer to our paper); your system may use either one (or neither!) during training.

搜集汇总

背景与挑战

背景概述

HookTheory数据集包含50小时的对齐旋律和声注释，源自HookTheory的TheoryTab DB，以JSON格式提供并采用CC BY-NC-SA 3.0许可证。该数据已预拆分为训练、验证和测试子集，适用于旋律转录模型训练，由斯坦福大学于2022年发布。