Multilingual i-Vector Based Statistical Modeling for Music Genre Classification

Abstract

For music signal processing, compared with the strategy which models each short-time frame independently, when the long-time features are considered, the time-series characteristics of the music signal can be better presented. As a typical kind of long-time modeling strategy, the identification vector (i-vector) uses statistical modeling to model the audio signal in the segment level. It can better capture the important elements of the music signal, and these important elements may benefit to the classification of music signal. In this paper, the ivector based statistical feature for music genre classification is explored. In addition to learn enough important elements for music signal, a new multilingual i-vector feature is proposed based on the multilingual model. The experimental results show that the multilingual i-vector based models can achieve better classification performances than conventional short-time modeling based methods.

A Multi-Modal Feature Embedding Approach to Diagnose Alzheimer Disease from Spoken Language

“…These two models are used to extract a feature vector from variable length audio signal. In addition to speaker identification tasks, these models are used in other areas such as language identification [25], emotion recognition [26], music genre classification [27], and online signature verification [28].…”

An Industry Driven Genre Classification Application using Natural Language Processing

“…As it can be seen the combination of the bass, vocals and accompaniment without velocities gives us the best performance. The best of these scores are either equal to better than the peak performance of [16].…”

“…The overall architecture of the system follows the process as described by it. Some of the ideas of extracting spectral features from the audio track itself, using a deep neural network and supplying it to a classifier, as explained in [16] have been adopted within this thesis extensively. Using inspiration from the aforementioned publication, thirty second clips from the start of every track are used as input into the proposed method.…”

“…Multilingual deep neural networks [16] are another avenue of research when trying to classify music genres. These have extensively been used in low resource automatic speech recognition (ASR).…”

“…When translated to the music information retrieval domain, this publication utilizes the nearest neighbor algorithm to find the most similar resource data from a large similar database. While [3] uses the Million Song Database (which is used by this thesis, this publication [16] trains on the GTZAN [5] and the ISMIR 2014 Genre Database [6]. The peak accuracy when using three layers each with 1024 hidden nodes, is stated too bee 93.4% according to this publication.…”