Statistical and Visual Analysis of Audio, Text, and Image Features for Multi-Modal Music Genre Recognition
Round 1
Reviewer 1 Report
The paper is interesting as the combination of audio features, album images, and lyrics is still rarely seen in the literature. It also fits the journal scope of interest.
Title: Statistical and Visual Analysis of Audio, Text, and Image
Features for - I am not sure whether the tile focuses on the main goal of the study; maybe it would be better to change its order “Multi-Modal Music Genre Recognition based Audio, Text, and Image
Features”?
After such a thorough analysis in Section 4 that covers a lot of material, Conclusions and Future Work are somewhat disappointing. I suggest including some general comments concerning the analyses performed. Moreover, “future work” is written in a style that does not communicate any useful information. I propose to rewrite the last paragraph of this Section and include more details. You may refer to problems and difficulties that should be further analyzed.
Did you evaluate how time-effective your method is? This issue should be addressed.
General comments:
The first (and to some extend 2nd) paragraph in Introduction seems obsolete. I suggest starting with the third one and then passing on the Related work.
The same remark concerns Section 3.2 – it covers well-known issues. At least, it may be shortened.
Although the list of References is quite long, the state-of-the-art is not complete (see comment below).
There are quite many language issues in the text. I have listed some of them beneath. Moreover, notably, punctuation and article should also be checked.
I propose to use References instead of Footsteps.
Starting a sentence with square brackets containing reference number is not very comprehensible. It is better to include the author’s name, i.e., “According to Sturm [1], ….
I am not sure about Appendices; maybe they should be placed under the Internet link?
At least, References should be before Appendices.
Detailed remarks:
Abstract:
mostly the audio -> mainly
to a pure learning -> to pure learning
a higher interpretability -> higher
and a futher -> and further
increase of -> in
Section 1
As features to represent music tracks, usually audio signal characteristics are calculated. -> not comprehensible
“independently of its popularity”??? - is this a necessary statement?
Section 2
“The use of album covers to determine music genres has been less investigated until now…” and “However, none of these approaches took albums covers into account.” -> There are a few recent literature sources that analyze relationship between album covers and music genres, they should be contained in References
we take also -> we also take
less parameters -> fewer
only few -> only a few
is is
meta data -> meta data
15 second -> 15-second
This sentence is not comprehensible: “Second, we estimate text features for lyrics and not album reviews; although in [2] it is argued that relevant genre information must not be captured in reviews and “will unlikely comply with the current taxonomy of the collection to be classified”, is is indeed probable that some reviews may contain such information and be compiled from many other mixed sources (experiences of a reviewer, meta data, artist information, etc.), so that lyrics can be rated as a more “pure” and “independent” modality.”
Section 3
a rigorous -> rigorous
by certain -by specific
can be always automatically -> can always be automatically
the Figure -> Figure
Figs. 2 and 3 – I am not sure whether “item” is an appropriate word here
A better measurement of classification quality -> evaluation?
cross validation -> cross-validation
A prominent example of such models are -> is
in phase domain -> in the phase domain
predicted also -> also predicted
the high comprehensibility -> high comprehensibility
possibly -> possible
a training set -> the
the final prediction is done -> made
Section 4
of the classifier types??? Naïve Bayes, SVM, and Random Forest -> types, i.e., …?
in the Tables 1 and 2 -> in Tables 1 and 2
features BoW -> BoW features
in the Tables 3 and 4 -> in Tables 3 and 4
of the image features BoF SIFT -> of the image BoF SIFT features
the second one image features, … -> the second one − image features, etc.
worth to note -> worth noting
any obvious -> any noticeable
Author Response
Please see the attachment
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The authors introduce a fusion technique to predict genres based on audio tracks, album cover images, and lyrics.
In the experimental section, the combination of the three modalities showed good performance in several cases.
Both content organization and experimentation are evaluated as well-organized.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Thank you for taking into account my comments. I appreciate that. In my opinion, the text is now much better organized and its aims are clearly presented.
