M3 (MADHAV Lab Mistake Detection for Music Teaching Database)
收藏NIAID Data Ecosystem2026-05-01 收录
下载链接:
https://zenodo.org/record/8332077
下载链接
链接失效反馈官方服务:
资源简介:
Problem Statement:
Music pedagogy, particularly in India, is an oral-aural tradition where the teacher sings along with the learner, mostly one-on-one. In the most common setup, teaching proceeds in a call-and-response fashion, where the teacher sings a short piece, and the learner repeats the same. For instance, the teacher may trim the original piece to include only the segment sung incorrectly by the learner. If the mistake persists or corresponds to an essential concept, the teacher may take a detour to a series of lessons aiming at cultivating that particular concept. This work is a step towards emulating this paradigm computationally. The teacher’s lessons are recorded priori. A learner is presented with these lessons, one piece at a time, sequentially. As the learner repeats the piece, her mistakes are to be detected computationally. Detection of singing mistakes is the primary task of this work. This is to be accomplished by comparing the teacher’s audio with that of the learner. The concept of tonic or keynote is very essential in Indian classical music. Generally, this note, often along with some other essential ones, is played throughout the performance via a drone instrument known as tanpur ¯ a¯. The pedagogy involves practicing with the drone instrument playing in the background.
Dataset:
The dataset thus prepared consists of audio files of two teachers and their disjoint set of learners. A file refers to an audio recording for a single lesson and may typically range between a few seconds to a minute in length. The mistakes of the learners are annotated by two expert teachers. They listen to the teacher’s and the learner’s audio and mark the mistakes in terms of the start time, end time, and mistake category. The mistake can fall into one of the following categories:
F: frequency error
A: amplitude error
P: pronunciation error
T: timing error
O: other error. This is followed by a short description of the error
For details on dataset curation and preparation or using it for research, please refer to the supplementing paper here.
创建时间:
2023-09-11



