See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/352780489

Mood based music recommendation system

Conference Paper · June 2021

CITATIONS READS

2 4,508

5 authors, including:

Ankita Mahadik Vijaya Bharathi Jagan

Don Bosco Institute of Technology Pillai College of Engineering
1 PUBLICATION 2 CITATIONS 4 PUBLICATIONS 8 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Vijaya Bharathi Jagan on 16 March 2022.

The user has requested enhancement of the downloaded file.

 Mood based music recommendation system
Janvi Patel
Ankita Mahadik Shambhavi Milgir
Department of Information Technology
Department of Information Technology Department of Information Technology
Don Bosco Institute of Technology
Don Bosco Institute of Technology Don Bosco Institute of Technology
Mumbai, India
Mumbai, India Mumbai, India
[email protected]
[email protected] [email protected]
Prof. Vijaya Bharathi Jagan Prof. Vaishali Kavathekar
Department of Information Technology Department of Information Technology
Don Bosco Institute of Technology Don Bosco Institute of Technology
Mumbai, India Mumbai, India
[email protected] [email protected]

Abstract— A user's emotion or mood can be detected by different expressions. Also, expressions of different or even
his/her facial expressions. These expressions can be the same people might vary for the same emotion, as
derived from the live feed via the system's camera. A lot emotions are hugely context dependent. While the focus can
of research is being conducted in the field of Computer on only those areas of the face which display a maximum of
Vision and Machine Learning (ML), where machines are emotions like around the mouth and eyes, how these
trained to identify various human emotions or moods. gestures are extracted and categorized is still an important
Machine Learning provides various techniques through question. Neural networks and machine learning have been
which human emotions can be detected. One such used for these tasks and have obtained good results.
technique is to use MobileNet model with Keras, which Machine learning algorithms have proven to be very useful
generates a small size trained model and makes in pattern recognition and classification, and hence can be
Android-ML integration easier. used for mood detection as well.
Music is a great connector. It unites us across With the development of digital music technology,
markets, ages, backgrounds, languages, preferences, the development of a personalized music recommendation
political leanings and income levels. Music players and system which recommends music for users is essential. It is
other streaming apps have a high demand as these apps a big challenge to provide recommendations from the large
can be used anytime, anywhere and can be combined data available on the internet. E-commerce giants like
with daily activities, travelling, sports, etc. With the Amazon, EBay provide personalized recommendations to
rapid development of mobile networks and digital users based on their taste and history while companies like
multimedia technologies, digital music has become the Spotify, Pandora use Machine Learning and Deep Learning
mainstream consumer content sought by many young techniques for providing appropriate recommendations.
people. There has been some work done on personalized music
People often use music as a means of mood recommendation to recommend songs based on the user’s
regulation, specifically to change a bad mood, increase preference. There exist two major approaches for the
energy level or reduce tension. Also, listening to the right personalized music recommendation. One is the content-
kind of music at the right time may improve mental based filtering approach which analyses the content of
health. Thus, human emotions have a strong relationship music that users liked in the past and recommends the music
with music. with relevant content. The main drawback of this approach
In our proposed system, a mood-based music player is that the model can only make recommendations based on
is created which performs real time mood detection and existing interests of the user. In other words, the model has
suggests songs as per detected mood. This becomes an limited ability to expand on the users' existing interests. The
additional feature to the traditional music player apps other approach is the collaborative filtering approach which
that come pre-installed in our mobile phones. An recommends music that a peer group of similar preference
important benefit of incorporating mood detection is liked. Both recommendation approaches are based on the
customer satisfaction. The objective of this system is to user’s preferences observed from the listening behaviour.
analyse the user’s image, predict the expression of the The major drawback of this approach is the popularity bias
user and suggest songs suitable to the detected mood. problem: popular (i.e., frequently rated) items get a lot of
exposure while less popular ones are under-represented in
Keywords—Face Recognition, Image Processing, Computer the recommendations. Generally, a hybrid approach is
Vision, Emotion Detection, Music, Mood detection implemented in which both content and collaborative
techniques are combined to extract maximum accuracy and
to overcome drawbacks of both types. [1]
I. INTRODUCTION
In this work, the aim is to create a music
Human emotions can be broadly classified as: fear, recommendation system/music player which will detect the
disgust, anger, surprise, sad, happy and neutral. A large user’s face, identify the current mood and then recommend a
number of other emotions such as cheerful (which is a playlist based on the detected mood.
variation of happy) and contempt (which is a variation of
disgust) can be categorized under this umbrella of emotions.
These emotions are very subtle. Facial muscle contortions
are very minimal, and detecting these differences can be
very challenging as even a small difference results in
 II. RELATED WORK music resources description. It is suggested that there is a
lack of systematic research on user behaviour and needs,
A. Literature Survey low level of feature extraction, and a single evaluation index
In a particular system [8], Anaconda and Python in current research. Situation was identified to be an
3.5 softwares were used to test the functionality and Viola- important factor in the music personalized recommendation
Jones and haar cascade algorithms were used for face system. Finally, it was concluded that when the weights
detection. Similarly, KDEF (Karolinska Directed Emotional given to all the contextual factors were the same, greatly
Faces) dataset and VGG (Visual Geometry Group) 16 were reduced the accuracy of the recommendation results.
used with CNN (Convolution Neural Network) model which Another research [12] states that their hybrid
was designed with an accuracy of 88%, for face recognition recommendation system approach concept will work once
and classification that validated the performance measures. their model is trained enough to recognize the labels. The
However, the results proved that the network architecture mechanism for the automatic management of the user
designed had better advancements than existing preferences in the personalized music recommendation
algorithms. Another system [9] used Python 2.7, Open- service automatically extracts the user preference data from
Source Computer Vision Library (OpenCV) & CK (Cohn- the user’s brain waves and audio features from music. In
Kanade) and CK+ (Extended Cohn-Kanade) database which their study, a very short feature vector, obtained from low
gave approximately 83% accuracy. Certain researchers have dimensional projection and already developed audio
described the Extended Cohn-Kanade (CK+) database for features, is used for music genre classification problems. A
those wanting to prototype and benchmark systems for distance metric learning algorithm was applied in order to
automatic facial expression detection. The popularity and reduce the dimensionality of the feature vector with a little
ease of access for the original Cohn-Kanade dataset this is performance degradation. Proposed user’s preference
seen as a very valuable addition to the already existing classifier achieved an overall accuracy of 81.07% in the
corpora. It was also stated that for a fully automatic system binary preference classification for the KETI AFA2000
to be robust for all expressions in a myriad of realistic music corpus. The user satisfaction was recognizable when
scenarios, more data is required. For this to occur very large brainwaves were used.
reliably coded datasets across a wide array of visual B. Existing Systems
variabilities are required (at least 5 to 10k examples for each
action) which would require a collaborative research effort • EMO Player: Emo player (an emotion-based
from various institutions. music player) is a novel approach that helps the
It was observed in a cross-database experiment [1] user to automatically play songs based on the
that raw features worked best with Logistic Regression for emotions of the user. [2]
testing RaFD (Radboud Faces Database) database and • SoundTree: Sound Tree is a music
Mobile images dataset. The accuracy achieved was 66% and recommendation system which can be integrated to
36% respectively for both using CK+ dataset as a training an external web application and deployed as a web
set. The additional features (distance and area) reduced the service. It uses people-to-people correlation based
accuracy of the experiment for SVM (Support Vector on the user's past behaviour such as previously
Machine) from 89%. The algorithm that had been listened, downloaded songs. [3]
implemented generalized the results from the training set to • lucyd: lucyd is a music recommendation tool
the testing set better than SVM and several other algorithms. developed by four graduate students in UC
An average accuracy of 86% was seen for RaFD database Berkeley's Master of Information and Data Science
and 87% for CK+ database for cross-validation=5.The main (MIDS) program. lucyd lets the user ask for music
focus was feature extraction and analysis of the machine recommendations using whichever terms they
algorithm on the dataset. But accurate face-detection want. [4]
algorithms become very important if there are multiple • Reel Time.AI: This system works by having the
people in the image. One of the works [10] was tested by user subscribe to them. The user can then upload
deriving expression from the live feed via the system's images of large gatherings such as shopping malls,
camera or any pre-existing image available in the memory. movie theatres and restaurants. The system then
It has been implemented using Python 2.7, OpenCV and identifies the moods happy and sad. It recognizes
NumPy. The objective was to develop a system that can which faces portray happy emotion and which
analyse the image and predict the expression of the person. faces portray sad emotion, and gives the verdict of
The study proved that this procedure is workable and the situation from the faces of the people present.
produces valid results. • Music.AI: It uses the list of moods as input for
There has also been research done on the Music mood of the user and suggests songs based on the
Recommendation System. According to one such research selected mood. It is a combination of Collaborative
[11], a preliminary approach to Hindi music mood filtering based and Content based filtering models.
classification has been described, that exploits simple Emotion, time, ambience and learning history are
features extracted from the audio. MIREX (Music the features taken into account for music
Information Retrieval Evaluation eXchange) mood recommendation. [5]
taxonomy gave an average accuracy of 51.56% using the
10-fold cross validation. In addition to this, there is an
article [10] that states that the current music
recommendation research results from the perspective of
C. Existing Algorithms/Tools • Mood Detection — Classification of the emotion
• Deep Learning based Facial Expression on the face as happy, angry, sad, neutral, surprise,
Recognition using Keras: Using this algorithm, up fear or disgust. For this task, the traditional Keras
to five distinct facial emotions can be detected in module of Python was used but, in the survey, it
real time. It runs on top of a Convolutional Neural was found that this approach takes a lot of time to
Network (CNN) that is built with the help of Keras train and validate and also works slowly when
whose backend is TensorFlow in Python. The integrated with android apps. So, MobileNet which
facial emotions that can be detected and classified is a CNN architecture model for Image
by this system are Happy, Sad, Anger, Surprise and Classification and Mobile Vision was used. There
Neutral. OpenCV is used for image processing are other models as well but what makes
tasks where a face is identified from a live webcam MobileNet special is that it has very less
feed which is then processed and fed into the computation power to run or apply transfer
trained neural network for emotion detection. Deep learning to. This makes it a perfect fit for Mobile
learning based facial expression recognition devices, embedded systems and computers without
techniques bring down to a greater extent, the GPU or low computational efficiency without
dependency on face-physics-based models and compromising the accuracy of the results. It uses
other pre-processing techniques by enabling depth wise separable convolutions to build light
lengthwise learning to occur in the pipeline directly weight deep neural networks. The dataset used for
from the input images. [14] training was obtained by combining FER 2013
• Hybrid approach of Music Recommendation: dataset [6] and MMA Facial Expression
There are several drawbacks to relying solely on Recognition dataset [7] from Kaggle. The FER
collaborative filtering to recommend music. The 2013 dataset contained grayscale images of size
biggest problem is the “Cold Start.” Music tracks 48x48 pixels. The MMA Facial Expression
are only tagged as often as listeners are discovering Recognition dataset had images of different
or listening to them. In other words, there are little specifications. Thus, all these images were
or no available ‘tags’ to describe new music or converted as per the images in FER 2013 dataset
music that has not been discovered yet. and combined to obtain an even larger dataset with
Additionally, listeners are more willing to supply 40,045 training images and 11,924 testing images.
tags for songs they enjoy most than for songs they MobileNet was used with Keras to train and test
mildly enjoy or do not enjoy at all. [13] our model for seven classes - happy, angry, neutral,
• Viola–Jones object detection framework: The sad, surprise, fear and disgust. We trained it for 25
Viola-Jones algorithm is a widely used mechanism epochs and achieved an accuracy of approximately
for object detection. The main property of this 75%.
algorithm is that training is slow, but detection is
fast. This algorithm uses Haar basis feature filters, B. Music Recommendation Module
so it does not use multiplications. The efficiency of The dataset of songs classified as per mood was found on
the Viola-Jones algorithm can be significantly Kaggle for two different languages - Hindi and English.
increased by first generating the integral Research for a good cloud storage platform to store, retrieve
image. [15] and query this song data as per user’s request was
conducted. Options like AWS, Google Cloud, etc. were
III. METHODOLOGY found but these were rejected as they were costly and
provided very limited storage for free. Then research for
The mood-based music recommendation system is an open-source streaming services like Restream.io, Ampache,
application that focuses on implementing real time mood etc. was conducted, but again, these services were web
detection. It is a prototype of a new product that comprises based/used for live streaming on YouTube/available only
two main modules: Facial expression recognition/mood for personal use. After a lot of research (and time
detection and Music recommendation. constraints), Firebase was considered a backend server. It
can be integrated with an android app just by one click and
A. Mood Detection Module its free plan provides storage of 5GB. But functions like
This Module is divided into two parts: user queries, server updates, etc. are a part of a paid plan so
it was decided to limit the scope of the project.
• Face Detection — Ability to detect the location of The mp3 versions of the songs were manually uploaded
face in any input image or frame. The output is the on Firebase storage and were linked in the Real Time
bounding box coordinates of the detected faces. For database as per mood and language (for filters).
this task, initially the python library OpenCV was
considered. But integrating it with an android app
C. Integration
was a complex task so the FaceDetector class
available in Java was considered. This library For the integration of these two modules in an Android
identifies the faces of people in a Bitmap graphic application, the trained MobileNet model was saved as an
object and returns the number of faces present in a .h5 file, and this .h5 file was then converted to a .tflite file
given image. using TensorFlow Lite Converter. It takes a TensorFlow
model as input and generates a TensorFlow Lite model as
output with .tflite extension. Since the MobileNet model is
used, the size of the tflite file is expected be around 20- 25
Megabyte (MB) which was the desired size. In Android
Studio, an assets folder was created to store the .tflite file
and labels.txt file. The labels.txt file contains the class labels
of the model. All the appropriate methods were created for
loading the model, running the interpreter and obtaining the
results.
A project on Firebase was created and mp3 songs were
uploaded in the storage section. These songs as per mood
and language in the real time database section. After this,
the Firebase database was linked to Android studio. An
appropriate UI for the android application was created and
the tflite model methods were linked with the songs on
Firebase. Finally, the application was tested to fix the bugs
if any.

Fig.1. displays the System Architecture Diagram and

Fig.2. displays the Data Flow Diagram of the system.

Fig.2. Data Flow Diagram of the system.

IV. HARDWARE AND SOFTWARE REQUIREMENTS

Fig.1. System Architecture Diagram.
A. Hardware Requirements
The system architecture diagram depicts the overall The most common set of requirements defined by any
outline of the software system and the relationships, operating system or software application is the physical
constraints, and boundaries between components. When the computer resources, also known as hardware. The hardware
user opens the android app, the main screen will be requirements required for this project are:
displayed which contains three buttons – take snap, use • Minimum 4 Gigabyte (GB) RAM (used for
emoji, play songs. If the user clicks on “take snap” button, processing)
the camera opens, user clicks picture. This picture is given
• Webcam (for testing on laptop/desktop)
as input to face detection program. If no face is detected or
• Minimum 16 Megapixel (MP) Resolution camera
multiple faces are detected, then an appropriate error
(for testing on android device)
message is displayed to the user. Upon successful single
face detection, the picture is given as input to the mood • 30 MB Memory space (approximate value)
detection module. The detected mood is displayed to the
user, after which the “play songs” button gets enabled. The B. Software Requirements
suitable playlist for the detected mood is displayed on the Software Requirements deal with defining software
playlist screen as shown in Fig.9 where the user can select resource requirements and prerequisites that need to be
and play the song. If the user presses the “use emoji” button, installed on a computer to provide optimal functioning of an
then a screen of five emojis will be displayed as shown in application. These requirements or pre-requisites are
Fig.10. User can click on any emoji to obtain the respective generally not included in the software installation package
playlist. To exit the app, the user has to just press the back and need to be installed separately before the software is
button.
installed. The software requirements that are required for
this project are:
• Python 3.6
• OpenCV 3.1
• PyCharm IDE
• Android Studio
V. RESULTS AND DISCUSSION
As every person has unique facial features, it is difficult
to detect accurate human emotion or mood. But with proper
facial expressions, it can be detected up to a certain extent.
The camera of the device should have a higher resolution.
The android application that we have developed runs
successfully and following are some of the screenshots
captured while using it. Fig.3. displays “sad” mood being
detected, Fig.4. displays “angry” mood being detected,
Fig.5. displays “happy” mood being detected, Fig.6.
displays “neutral” mood being detected and Fig.7. displays Fig.5. “Happy” mood detected successfully by the application.
“surprise” mood being detected. and Fig.8. displays “fear”
mood being detected.

Fig.6 “Neutral” mood detected successfully by the application.

Fig.3. “Sad” mood detected successfully by the application.

Fig.4. “Angry” mood detected successfully by the application. Fig.7. “Surprise” mood detected successfully by the application.
 picture of their face, which in turn is unable to detect the
proper mood, or any other reason, the user can click on the
“Use Emoji” button and select the emoji which represents
the mood which they are in, or the mood that they want their
playlist to be generated of. Fig.10. is the screenshot of the
screen that is displayed to the user when they click the “Use
Emoji” button. The first is the emoji for the mood “happy”,
the second for the mood “angry”, the third for the mood
“surprise”, the fourth for the mood “sad”, and the fifth for
the mood “neutral”.

Fig.8. “Fear” mood detected successfully by the application.

The system detects the mood in real time and a playlist is

displayed for that mood accurately. The playlist for “happy”
mood can be seen in Fig.9. It is able to fetch and play the
songs from the recommended playlist in the presence of a
stable internet connection. The playlist displays the songs to
the user in list view, and when the user selects a song to be Fig.10. The screen for the user to generate playlist via emojis.
played, the song is played in a media player, that comes
with all the necessary buttons such as play, pause, shuffle, In Fig.11(a), the graph displays the accuracy of our
next and previous, along with a seek bar. For “angry”, model, where the x-axis specifies the number of epochs and
“Fear”, “disgust” and “surprise” moods, devotional, the y-axis specifies the accuracy. As it can be seen in the
motivational and patriotic songs are suggested to the user. figure, our model has achieved approximately 75%
Hence, the user is also provided with mood improvement. accuracy. Since it is a fully computer-based system, it
understands emotions in the way it has been trained.

Fig.9. Playlist for “Happy” mood where a song is being played.

Fig.11(a). Training and Validation Accuracy.
We have provided our users with the additional option of
using emojis to generate the playlist. Whenever the user In Fig.11(b), the graph displays the training and
does not wish to, or is unable to take a snapshot of their validation loss of our model, where the x-axis specifies the
mood due to various reasons such as extremely high or low number of epochs and the y-axis specifies the loss.
lighting, their camera not working properly, they have a
lower resolution camera which is unable to take a clear
 example, the “disgust” mood is mostly classified as “angry”
mood since the facial features (eyebrows, cheeks) are
similar for both.
Thus, for more accurate results it needs to be trained for
more images, and for a greater number of epochs.
Recommendation of movies and TV series on the basis of
mood detection can also be considered as a future scope for
our project.
ACKNOWLEDGMENT
We would like to thank our project guides Prof. Vijaya
J. and Prof. Vaishali K. for the valuable support and
guidance they gave us on every step of the project execution.
We would also like to thank the project review committee
members Prof. Amiya T., Prof. Phiroj S. and Prof. Sunantha
K. We would also like to express our gratitude to Prof.
Janhavi B., Head of Department, Department of Information
Technology, Don Bosco Institute of Technology, Mumbai
who helped us accomplish this work.
Fig.11(b). Training and Validation Loss.

The following table provides a comparison with

respect to accuracy between the different existing systems
and tools and our proposed system. REFERENCES
[1] Raut, Nitisha, "Facial Emotion Recognition Using Machine Learning"
TABLE I. ACCURACY COMPARISON (2018). Master's Projects. 632. https://doi.org/10.31979/etd.w5fs-
s8wd
Sr. Accuracy
Existing Systems and Tools [2] Hemanth P,Adarsh ,Aswani C.B, Ajith P, Veena A Kumar, “EMO
No. (%) PLAYER: Emotion Based Music Player”, International Research
Deep learning-based face recognition Journal of Engineering and Technology (IRJET), vol. 5, no. 4, April
1. 88
using keras 2018, pp. 4822-87.
Hybrid approach for music
2.
recommendation
80.7 [3] Music Recommendation System: “Sound Tree”, Dcengo Unchained:
Sıla KAYA, BSc.; Duygu KABAKCI, BSc.; Işınsu KATIRCIOĞLU,
Viola Jones object detection
3. 86 BSc. and Koray KOCAKAYA BSc. Assistant : Dilek Önal
framework
Supervisors: Prof. Dr. İsmail Hakkı Toroslu, Prof. Dr. Veysi İşler
4. Music.AI 78.84 Sponsor Company: ARGEDOR
Mood based music recommendation [4] Tim Spittle, lucyd, GitHub, , April 16, 2020. Accessed on:
5. 75 [Online], Available at: https://github.com/timspit/lucyd
system
[5] A. Abdul, J. Chen, H.-Y. Liao, and S.-H. Chang, “An Emotion-Aware
Personalized Music Recommendation System Using a Convolutional
Neural Networks Approach,” Applied Sciences, vol. 8, no. 7, p. 1103,
Jul. 2018.
VI. CONCLUSION
[6] Manas Sambare, FER2013 Dataset, Kaggle, July 19, 2020. Accessed
Even though human emotions are complex and subtle, it on: September 9, 2020. [Online], Avalilable at:
is possible for a machine learning model to be trained to https://www.kaggle.com/msambare/fer2013
accurately detect a set of emotions which can be [7] MahmoudiMA, MMA Facial Expression Dataset, Kaggle, June 6,
2020. Accessed on: September 15, 2020. [Online], Available at:
differentiated from each other with certain facial https://www.kaggle.com/mahmoudima/mma-facial-expression
expressions. The expression on a person’s face can be used [8] Dr. Shaik Asif Hussain and Ahlam Salim Abdallah Al Balushi, “A
to detect their mood, and once a certain mood has been real time face emotion classification and recognition using deep
detected, music suitable for the person’s detected mood can learning model”, 2020 Journal. of Phys.: Conf. Ser. 1432 012087
be suggested. Our model, having the accuracy of [9] P. Lucey, J. F. Cohn, T. Kanade, J. Saragih, Z. Ambadar and I.
approximately 75%, is able to detect seven moods Matthews, "The Extended Cohn-Kanade Dataset (CK+): A complete
dataset for action unit and emotion-specified expression," 2010 IEEE
accurately: anger, disgust, fear, happy, sad, surprise and Computer Society Conference on Computer Vision and Pattern
neutral; and our android application is able to play the music Recognition - Workshops, San Francisco, CA, USA, 2010, pp. 94-
that would be suitable for the detected mood. 101, doi: 10.1109/CVPRW.2010.5543262.
For accurate detection of fear and disgust moods, [10] Puri, Raghav & Gupta, Archit & Sikri, Manas & Tiwari, Mohit &
Pathak, Nitish & Goel, Shivendra. (2020). Emotion Detection using
additional parameters such as heart rate or body temperature Image Processing in Python.
must also be considered rather than solely depending on [11] Patra, Braja & Das, Dipankar & Bandyopadhyay, Sivaji. (2013).
facial expressions. In addition to that, finding suitable music Automatic Music Mood Classification of Hindi Songs.
to be played on detection of fear or disgust mood is also a [12] Lee, J., Yoon, K., Jang, D., Jang, S., Shin, S., & Kim, J. (2018).
challenge. As a result, it can be considered as a future scope MUSIC RECOMMENDATION SYSTEM BASED ON GENRE
for our project. Our trained model is an overfit model, which DISTANCE AND USER PREFERENCE CLASSIFICATION.
can sometimes lead to fluctuations in accurate detection. For
[13] Kaufman Jaime C., University of North Florida, “A Hybrid Music Tags and Acoustic Features”, UNF Digital Commons, 2014.
Approach to Music Recommendation: Exploiting Collaborative
[14] D Priya, Face Detection, Recognition and Emotion Detection in 8
lines of code!, towards data science, April 3, 2019. Accessed on: July
12, 2020 [Online], Available at: https://towardsdatascience.com/face-
detection-recognition-and-emotion-detection-in-8-lines-of-code-
b2ce32d4d5de

[15] bluepi, “\Classifying Different Types of Recommender Systems,

November 14, 2015. Accessed on: July 7, 2020. [Online],
Available on:https://www.bluepiit.com/blog/classifying-
recommender-
systems/#:~:text=There%20are%20majorly%20six%20types,syste
m%20and%20Hybrid%20recommender%20system

View publication stats