MERGE App: A Prototype Software for Multi-User Emotion-Aware Music Management

MERGE App: A P o o ype So wa e o Mul i-use Emo ion-awa e Music Managemen Ped o Lima Lou o1 a, Guilhe me B anco1 b, Hugo Redinho1 c, Rica do Co eia1 d, Rica do Malhei o1,2 e, Rena o Panda1,3 , Rui Ped o Pai a1 g 1Uni e si y o Coimb a, Cen e o In o ma ics and Sys ems o he Uni e si y o Coimb a (CISUC), Depa men o In o ma ics Enginee ing, and LASI 2Poly echnic Ins i u e o Lei ia School o Technology and Managemen 3Ci2 — Sma Ci ies Resea ch Cen e , Poly echnic Ins i u e o Toma ped olou [email p o ec ed], guilhe me[email p o ec ed], [email p o ec ed], { ica doco eia, smal, panda, uiped o}@dei.uc.p Keywo ds: Music In o ma ion Re ie al; Music Emo ion Recogni ion; Machine Lea ning; Deep Lea ning; So wa e Abs ac : We p esen a p o o ype so wa e o mul i-use music lib a y managemen using he pe cei ed emo ional con en o songs. The ool o e s music playback ea u es, song il e ing by me ada a, and au oma ic emo ion p edic ion based on a ousal and alence, wi h he possibili y o pe sonalizing he p edic ions by allowing each use o edi hese alues based on hei own emo ion assessmen . This is an impo an ea u e o handling bo h classi ica ion e o s and subjec i i y issues, which a e inhe en aspec s o emo ion pe cep ion. A pa h- based playlis gene a ion unc ion is also implemen ed. A mul i-modal audio-ly ics eg ession me hodology is p oposed o emo ion p edic ion, wi h accompanying alida ion expe imen s on he MERGE da ase . The esul s ob ained a e p omising, showing highe o e all pe o mance on ain- alida e- es spli s (73.20% F1- sco e wi h he bes da ase /spli combina ion). 1 INTRODUCTION The digi al e a has b ough an unp eceden ed amoun o music igh a ou inge ips h ough digi al ma - ke places and s eaming se ices. Wi h he sudden a ailabili y o millions o songs o use s, he neces- si y o au oma ically o ganize and ind ele an mu- sic eme ged. Cu en ecommenda ion sys ems p o- ide pe sonalized sugges ions o use s based on lis- ening pa e ns and using ags, such as gen e, s yle, e c. Howe e , op ions a e lacking when we conside ecommenda ions based on he au oma ic analysis o he emo ional con en o songs. The ield o Music Emo ion Recogni ion (MER) has seen conside able ad ances in ecen yea s in e ms o he mo e classical app oaches. Panda e al. ah ps://o cid.o g/0000-0003-3201-6990 bh ps://o cid.o g/0000-0003-4073-1716 ch ps://o cid.o g/0009-0004-1547-2251 dh ps://o cid.o g/0000-0001-5663-7228 eh ps://o cid.o g/0000-0002-3010-2732 h ps://o cid.o g/0000-0003-2539-5590 gh ps://o cid.o g/0000-0003-3215-3960 (2020) p oposed a new se o ea u es ha consid- e ably inc eased he pe o mance o hese sys ems, achie ing a 76.4% F1-sco e wi h he op 100 anked ea u es. Al hough he ea u e e alua ion is limi ed o one da ase , he imp o emen s a e signi ican com- pa ed o he bes esul s om simila sys ems ha eached a glass ceiling Hu e al. (2008). One d awback o audio-only me hodologies is hei sho comings when di e en ia ing alence. Va - ious sys ems ha e been p oposed using a bimodal app oach le e aging bo h audio and ly ics, a aining conside able imp o emen s when compa ed o sys- ems using only one o he o he Delbouys e al. (2018); Py o olakis e al. (2022). Such sys ems ha e also implemen ed Deep Lea ning (DL) a chi ec u es o skip he ime-consuming ea u e enginee ing and ex ac ion s eps om he classical sys ems and con- side ably speed up he in e ence p ocess o he o e all sys em. In his s udy, we p esen he MERGE1applica ion, 1MERGE is he ac onym o ”Music Emo ion Recog- ni ion nEx Gene a ion”, a esea ch p ojec unded by he Po uguese Science Founda ion. Figu e 1: MERGE applica ion in e ace. The AV plo , alongside music playback and song display con ols, is seen in yellow. The able iew is highligh ed in blue wi h ed highligh il e ing sea ch ba and bu on o adding music. Finally, g een highligh s he bu ons o applica ion in o ma ion and use logou . which au oma ically p edic s he a ousal and alence o songs based on Russell’s Ci cumplex Model Rus- sell (1980). Two axes make up his model: a ousal (Y-axis), which depic s whe he he song has high o low ene gy, and alence (X-axis), which ep esen s whe he he emo ion o he song has a nega i e o pos- i i e conno a ion. The in eg a ed model used o p edic ion is also p esen ed in his s udy alongside alida ion expe i- men s, which ecei ed bo h audio and ly ics in o ma- ion o map he song mo e accu a ely in o he abo e men ioned model. The MERGE applica ion is a ollow-up o he MOODe ec o applica ion, p e iously c ea ed by ou eam Ca doso e al. (2011). The new MERGE app was buil om sc a ch, wi h signi ican code e ac o - ing and op imiza ion, while keeping he o e all use in e ace o he MOODe ec o app. In addi ion, sig- ni ican no el ea u es and imp o emen s we e im- plemen ed, namely: i) a bimodal app, which ex- ploi s he combina ion o audio and ly ics da a o im- p o ed classi ica ion (unlike he single audio modali y in he MOODe ec o app); ii) an imp o ed classi ica- ion model, aining wi h he MERGE da ase Lou o e al. (2024b) and ollowing a deep lea ning app oach Lou o e al. (2024a); iii) and a shi om he mono- li hic single-use pa adigm o he web-based mul i- use pa adigm. 2 MERGE APPLICATION The MERGE applica ion is implemen ed using Ja aSc ip , wi h he addi ion o he jQue y lib a y o handle AJAX, o i s on end, while he backend is se ed using he Exp ess lib a y on op o Node.js. The applica ion in e ace is depic ed in Figu e 1. 2.1 Applica ion O e iew The componen s can be b oken down as: i) he Rus- sell’s Ci cumplex model whe e all songs can be seen (highligh ed in yellow); ii) a able iew o he songs wi h a ious op ions o so ing (highligh ed in blue); iii) op ions o il e and add new songs (seen in he ed egion, om le o igh ); i ) in o ma ion abou he applica ion, he cu en use , and an op ion o lo- gou (highligh ed in g een, also om le o igh ). Songs a e placed in he plo desc ibed in i) acco d- ing o he es ima ed a ousal and alence (AV) alues in an in e al o [-1, 1] o each axis. The p ocess o ob aining hese alues is desc ibed in Sec ion 3. Beyond he AV posi ioning, each poin on he plane is also colo -coded depending on he quad an : g een (happy), ed ( ense), blue (sad), and yellow ( elaxed). The iew o he g aph can be swi ched be ween a) ”Uploaded” o show only songs uploaded by he cu - en use , b) ”My Lib a y” o display a use ’s lib a y, i.e., he songs uploaded by he cu en use , plus songs uploaded by o he use s added by he cu en use , and c) ”All songs” o show all he songs a ailable in he da abase. A no e ega ding he la e op ion is he di - e en ia ion o songs no added by he use , appea ing as g ey do s in he plo , also depic ed in Figu e 1. Each use can change he song’s posi ion di ec ly by mo ing he poin in he plane iew, o by edi ing he AV alues h ough he able iew. These new AV alues a e unique o he use . The applica ion can be used as an audio playback so wa e, hus o e ing usual ea u es such as: • Playback con ols o mp3 iles (play, pause, seek); • Volume con ols, including mu e; • Double-clicking a song o be played ei he in he plo o able iew; Figu e 2: MERGE App Backo ice. Use s wi h adminis a i e p i ileges can o e w i e he model used o AV alues’ p edic- ion (highligh ed in g een) and expo a CSV ile wi h in o ma ion ega ding he anno a ions o all use s o each song in he da abase (highligh ed in blue). Figu e 3: En i y ela ion diag am o he applica ion’s da abase. • Fil e ing and so ing by any o he a ailable song p ope ies ( i le, a is , alence, a ousal, emo ion); • Adding and dele ing songs om he use ’s lib a y. The applica ion also p o ides a backo ice o use s wi h adminis a i e p i ileges, pic u ed in Fig- u e 2. A e logging in, he use can pe o m one o wo ac ions: upload and deploy a new model o AV p edic ion, and expo a CSV wi h he exis ing use AV anno a ions. The la e op ion is designed o eas- ily e ie e each use ’s a ailable anno a ions o songs in hei espec i e lib a ies. In his way, he MERGE app can be used as a c owdsou cing da a collec ion and anno a ion ool, p omo ing he c ea ion o size- able and quali y MER da ase s, a cu en key need in MER esea ch Panda e al. (2020). Rega ding he da abase used o s o e all he el- e an da a om use s and songs, he co esponding en i y ela ionship diag am is depic ed in Figu e 3. The use able s o es he use ’s pe sonal in o ma ion, as well as he use ole, o access p i ileges pu - poses. The pa h o he model used o AV p edic- ion is sa ed in he co esponding able and iden i ies he use ha uploaded he cu en ly deployed model. The song able s o es all song- ela ed in o ma ion, in- cluding me ada a, he AV alues i s p edic ed by he p esen ly deployed model, he mapped quad an in he plo iew, he pa h o he uploaded audio clip, and he e e ence o he use who i s added he song. The use -speci ic anno a ions a e s o ed in he an- no a ion able, which s o es he AV alues de ined pe he use ’s pe cep ion, he co esponding quad an , and a e e ence o he song and use o ha anno a ion. Finally, he lib a y able s o es all use lib a ies, con- aining only songs added by he use , ei he h ough uploading o om o he use s’ lib a ies. Figu e 4: A he op, ”Ano he One Bi es The Dus ” by Queen is added o he lib a y and placed in he plane acco ding o he p edic ed AV alues. A he bo om, he poin ep esen ing he song is mo ed o a mo e accu a e posi ion, acco ding o he use . 2.2 Building An Emo ionally-awa e Lib a y A e adding a new song om an a ailable MP3 ile, AV alues a e au oma ically p edic ed, and a poin is added o he plo alongside a new en y on he able. Should he use disag ee wi h he p edic ed alues, hese can be easily changed by edi ing he en y om he able iew o mo ing he poin in he plo . An ex- ample o he ini ially p edic ed posi ion o a newly added song and he inal posi ion a e adjus men can be seen in Figu e 4. This pe sonaliza ion mechanism p o ides use s wi h he abili y o add ess he in in- sic subjec i i y in MER. Howe e , ackling his issue con inues o p esen a signi ican challenge. 2.3 Pa h-based Playlis Gene a ion The abili y o gene a e a playlis based on a use - d awn pa h is cu en ly implemen ed, as depic ed in Figu e 5. This ea u e allows use s o eely c ea e an emo ionally- a ying playlis . This is done by compu ing he dis ance o he use - de ined Ncloses songs o he e e ence poin s ha make up he d awn pa h. The use may also con igu e how a he songs can be om he pa h o be consid- e ed in o he calcula ions. This h eshold is de ined as a decimal numbe be ween he plane in e al ([-1, 1]). 3 SONG EMOTION PREDICTION In his sec ion, we discuss he me hodology used o p edic AV alues o a gi en song. Fi s , he DL model’s a chi ec u e is p esen ed, ollowed by he p e- p ocessing s eps o each modali y, a desc ip ion o he op imiza ion used, and he e alua ion conduc ed. 3.1 Model A chi ec u e The p oposed a chi ec u e, depic ed in Figu e 6, is based on he one by Delbouys e al. Delbouys e al. (2018). Dis inc audio and ly ics b anches ecei e Mel-spec og am ep esen a ions and wo d embed- dings, espec i ely. The lea ned ea u es o each modali y a e hen used and u he p ocessed by a small Dense Neu al Ne wo k (DNN), inally ou - pu ing he AV alues p edic ion. We op ed o a bimodal audio-ly ics app oach conside ing ha bo h modali ies ha e ele an in o - ma ion o he di e en axes o Russell’s Ci cum- plex Model. Audio has been shown o be e p edic Figu e 5: The use can be seen d awing a pa h o gene a e a playlis wi h he desi ed emo ional ajec o y a he op. The esul o he pa h-based playlis gene a ion is p esen ed a he bo om. a ousal, while ly ical in o ma ion is mo e ele an o alence p edic ion Lou o e al. (2024b). S a ing in he audio b anch, Mel-spec og am ep esen a ions o each sample a e ed o he ea u e lea ning po ion o he baseline a chi ec u e p esen ed in Lou o e al. Lou o e al. (2024a). I is composed o ou con olu ional blocks, composed o a 2D Con- olu ional laye , ollowed by a Ba ch No maliza ion, D opou , and Max Pooling laye , inishing wi h ReLU ac i a ion. As o he ly ics b anch, he wo d em- beddings o ly ics a e also ed o ou con olu ional blocks, each comp ising a 1D Con olu ional laye , ollowed by Max Pooling and a ReLU ac i a ion lay- e s. To balance he in o ma ion om each modali y, we signi ican ly educe he o e whelming amoun o lea ned ea u es om ly ics using a Dense laye be- o e me ging he lea ned ea u es om bo h b anches. The classi ica ion po ion o he model is com- posed o al e na ing D oupou and Dense laye s, which educe and u he p ocess he se o ea u es espec i ely, inally ou pu ing one o Russell’s Ci - cumplex model’s ou quad an s. 3.2 P e-p ocessing S eps A se o p e-p ocessing s eps is necessa y o ob ain he da a ep esen a ions used o each b anch o he a chi ec u e de ailed abo e. The lib osa lib a y McFee e al. (2015) is used o ob ain he Mel-spec og am ep esen a ion o he au- dio b anch. The audio samples, p o ided as mp3 iles, a e i s con e ed o wa e o ms (.wa ) and down- sampled om 22.5 o 16kHz. This is done o educe he complexi y o he model, along wi h he compu- a ional cos o op imiza ion. The downsampling has been shown o p o ide simila esul s o highe sam- pling a es, showing he obus ness o DL app oaches Py o olakis e al. (2022). The spec al ep esen a ions a e hen gene a ed using de aul pa ame e s o he leng h o he Fas Fou ie T ans o m window (2048) as well as he hop size (512). As o wo d embeddings, he Sen ence T ans- o me lib a y om Hugging Face was used, speci i- cally, he all- obe a-la ge- 1 p e- ained model. The embedde ecei es a con ex o up o 512 okens and ou pu s a 1024 embedded ec o . Gi en ha he bes esul s we e p o ided by using he ull con ex win- dow, some o he ly ics had o be cu o a some poin . A e some simple okeniza ion s eps, namely emo ing new line cha ac e s and con e ing all ex o lowe case, he embeddings we e ob ained up o he al eady men ioned con ex size. Figu e 6: The mul i-modal audio-ly ics eg ession model. Emo ionally- ele an ea u es a e lea ned o bo h he audio ep e- sen a ion in he Mel-spec og am- ecei ing b anch and he ly ics ep esen a ion in he b anch ecei ing he p e iously gene - a ed wo d embeddings. AV alues a e p edic ed a e conca ena ing and p ocessing he lea ned ea u es om bo h b anches. 3.3 Model Op imiza ion Model op imiza ion was conduc ed using he Bayesian op imiza ion implemen a ion o he Ke as Tune lib a y O’Malley e al. (2019). This me hod inds he bes combina ion o hype pa ame e s in p e- iously de ined in e als o each, ei he maximizing o minimizing an objec i e unc ion de ined by he use . Since ou me hodology is based on a eg ession ask o p edic a ousal and alence o a gi en sam- ple, he objec i e is de ined as minimizing he sum o he mean squa ed e o (MSE) o bo h. This en- su es ha none is p io i ized, le e aging bo h audio’s be e p edic abili y in e ms o a ousal and he same o ly ics’ p edic abili y o alence. The in e als o each conside ed hype pa ame e , namely, ba ch size, op imize , and co esponding lea ning a e, a e p e- sen ed in Table 1. Table 1: Op imal Hype pa ame e s Fo Each Da ase Bes Hype pa ame e s Ba ch Size Op mize Lea ning Ra e 64 SGD 1e-2 The op imiza ion p ocess is un o e en ials, pe he lib a y’s de aul , s a ing a he lowe end o each in e al. Fo each ial, he model is ained o a max- imum o 200 epochs, wi h an ea ly s opping s a egy de ined o check o no imp o emen s o he alida- ion loss o 15 consecu i e epochs. This conside ably educes he ime needed o conduc he ull op imiza- ion phase since less ime is spen on unde pe o ming se s o hype pa ame e s. We used a 70-15-15 ain- alida e- es (TVT) spli as ou alida ion s a egy, as de ined in Lou o e al. (2024b). The esul ing models o each ial a e backed up o la e usage, including he e alua ion phase, which is discussed nex . 3.4 Da a and E alua ion The MERGE Bimodal Comple e da ase was used o alida ing ou app oach. P oposed in Lou o e al. (2024b), i comp ises a se o 2216 bimodal samples (audio clips and co esponding ly ics). Fo each sam- ple, he da ase p o ides a 30-second audio exce p o he mos ep esen a i e pa o he song, links o he ull ly ics, labels co esponding o each o he quad- an s in Russell’s Ci cumplex model, and AV alues, used o ob ain he p e iously men ioned labels, cal- Table 2: TVT 70-15-15 Resul s Fo MERGE Audio Comple e F1-sco e P ecision Recall R2 RMSE (A/V) (A/V) 73.20% 74.53% 73.49% 0.454 0.133 0.506 0.339 cula ed based on he ex ac ed emo ion- ela ed ags a ailable in AllMusic 2. The abo e-men ioned AV alues a e ob ained h ough he ollowing p ocess. Fi s , he a ailable ags o each song in he da ase a e ob ained om he All Music pla o m. Using Wa ine ’s Adjec i e Dic iona y Wa ine e al. (2013), he exis ing ags a e ansla ed o a ousal and alence alues. Finally, The alues a e hen a e aged ac oss all ags co espond- ing o a speci ic song, ob aining i s inal mapping on Russell’s Ci cumplex model. Fo he TVT s a egy, bo h he aining and ali- da ion se s a e used in he op imiza ion unc ion. The se o op imal hype pa ame e s is ound using he la - e . A e aining he model o each da ase , he ol- lowing me ics a e compu ed be ween he ac ual and p edic ed AV alues in he es se o each class as well as o he o e all pe o mance: F1-sco e, P eci- sion, Recall, R2(squa ed Pea son’s co ela ion), and Roo Mean Squa ed E o (RMSE). Be o e compu ing hese me ics, he p edic ed and eal AV alues we e mapped o Russell’s Ci cumplex model o ob ain classes o calcula ing P ecision, Re- call, and F1-sco e. 4 EXPERIMENTAL RESULTS AND DISCUSSION Tables 2 and 3 show he o e all esul s o he dis- cussed me hodology. The a ousal and alence s an- dalone esul s o he R2and RMSE me ics a e p e- sen ed in consecu i e lines in he o de displayed in he ables. The ob ained esul s o bo h da ase s a e lowe han hose ob ained in p e ious s udies ocused on s a ic MER as a ca ego ical p oblem Lou o e al. (2024a). The bes esul a ained is a 73.20% F1- sco e, which is a ound 6% lowe han he esul s ob- ained o he same da ase and e alua ion s a egy in he men ioned a icle. The lowe esul s a e mos ly due o he semi-au oma ic app oach o ob ain AV al- ues (see Sec ion 3.4, conside ing ha he ags a ail- able on All Music a e use -gene a ed and i s cu a ion is unknown. 2h ps://www.allmusic.com/ Table 3: TVT 70-15-15 Resul s Con usion Ma ix Fo MERGE Audio Comple e P edic ed Q1 Q2 Q3 Q4 Ac ual Q1 61.3% 10.4% 6.6% 21.7% Q2 9.8% 82.4% 5.9% 2.0% Q3 1.4% 4.3% 78.3% 15.9% Q4 7.3% 0.0% 18.2% 74.5% As shown in Table 2, he R2me ic o alence ou pe o med he one o a ousal, al hough ha ing a la ge RMSE. This indica es ha he ela i e alence h oughou songs is easonably cap u ed, despi e he la ge RMSE e o . Al hough he a ained esul s show oom o im- p o emen , hey a e a good s a ing poin o he use . Gi en he subjec i e na u e o each use ’s emo ional pe cep ion, we belie e ha he pe sonaliza ion ea u e included in he MERGE app is a aluable mechanism o handling subjec i i y in MER. In e ms o he esul s o sepa a e quad an s (Ta- ble 3), we can see ha some Q1 songs a e con used wi h Q4 songs (21.77% Q1 songs a e inco ec ly clas- si ied as Q4). Mo eo e , he e is also some con usion be ween Q3 and Q4 (15.9% o Q3 songs a e p edic ed as Q4 and 18.2% o Q4 songs a e classi ied as Q3). This is a known di icul y in MER, as discussed in Panda e al. (2020) ha needs u he esea ch. 5 CONCLUSION AND FUTURE WORK We p esen ed he p o o ype o he MERGE applica- ion. Cu en ly, he ini ial e sion has implemen ed music playback ea u es, he abili y o add and il e songs o a sha ed da abase, lis and plane iews, he la e based on Russell’s Ci cumplex model, and use managemen unc ionali ies. Mo eo e , a bimodal audio-ly ics model is inco po a ed in o he backend o he p o o ype o allow o AV alue p edic ion o use -uploaded songs. Pa h-based playlis gene a ion has also been implemen ed, enabling use s o c a a playlis ha ollows a speci ic emo ional ajec o y hey ha e selec ed. S ill, many mo e unc ionali ies a e planned o he applica ion in u u e i e a ions. The highligh ed unc ionali ies include use -gene a ed ags o a mo e cus omized il e ing expe ience ha would be a ail- able o o he use s; au oma ic ly ics o he ull song sc aped om an a ailable API, e.g., Genius; and Mu- sic Emo ion Va ia ion De ec ion (MEVD) p edic ion suppo , including isualiza ion wi h he same colo code used in he plo . A s andalone desk op applica- ion is also planned wi hou he c oss-use ea u es. in addi ion o implemen ing hese upcoming ea u es, We plan o conduc in-dep h use expe ience s udies o gain a mo e comp ehensi e unde s anding o he sys em’s e icacy and use sa is ac ion. Valida ion expe imen s on wo ecen ly p oposed da ase s a e p o ided alongside a ho ough sys em de- sc ip ion, elaying insigh s in o he ob ained esul s. These a e s ill below he ca ego ical app oach p e- sen ed in Lou o e al. (2024b) due o he al eady dis- cussed semi-au oma ic AV mapping app oach in Sec- ion 3.4. Despi e his, he p edic ions a e a good s a - ing poin o be u he adjus ed o he use ’s pe cep- ion. Rega ding he ac ual model, nei he ea u e lea n- ing po ion may be ideal o he p oblem a hand since hey we e o iginally de eloped o a ca ego i- cal p oblem. De eloping mo e sui able a chi ec u es should hus be conside ed u u e wo k. Fu he mo e, he da a ep esen a ions, especially he wo d embed- dings, may also be u he imp o ed, conside ing ha he p e- ained model used is limi ed o a con ex win- dow o 512 okens. To conclude, we belie e he p oposed app migh be use ul o music lis ene s. Al hough he e is oom o imp o emen (as he a ained classi ica ion esul s show), he pe sonaliza ion mechanism is a use ul ea- u e o handling p edic ion e o s and subjec i i y. Finally, he pe sonaliza ion ea u e and he mul i-use en i onmen ha e he po en ial o acqui e quali y use anno a ions, leading o a u u e la ge and mo e obus MER da ase . ACKNOWLEDGEMENTS This wo k is unded by FCT - Founda ion o Sci- ence and Technology, I.P., wi hin he scope o he p ojec s: MERGE - DOI: 10.54499/PTDC/CCI- COM/3171/2021 inanced wi h na ional unds (PID- DAC) ia he Po uguese S a e Budge ; and p ojec CISUC - UID/CEC/00326/2020 wi h unds om he Eu opean Social Fund, h ough he Regional Ope a- ional P og am Cen o 2020. Rena o Panda was sup- po ed by Ci2 - FCT UIDP/05567/2020. We hank all e iewe s o hei aluable sugges- ions, which help o imp o e he a icle. REFERENCES Ca doso, L., Panda, R., and Pai a, R. P. (2011). Moode ec- o : A p o o ype so wa e ool o mood-based playlis gene a ion. In Simp´ osio de In o m´ a ica - INFo um 2011, Coimb a, Po ugal. Delbouys, R., Hennequin, R., Piccoli, F., Royo-Le elie , J., and Moussallam, M. (2018). Music Mood De ec ion Based On Audio And Ly ics Wi h Deep Neu al Ne . In P oceedings o he 19 h In e na ional Socie y o Music In o ma ion Re ie al Con e ence, pages 370–375, Pa is, F ance. Hu, X., Downie, J. S., Lau ie , C., Bay, M., and Ehmann, A. F. (2008). The 2007 Mi ex Audio Mood Classi ica ion Task: Lessons Lea ned. In P oceedings o he 9 h In e - na ional Socie y o Music In o ma ion Re ie al Con- e ence, pages 462–467, D exel Uni e si y, Philadelphia, Pennsyl ania, USA. Lou o, P. L., Redinho, H., Malhei o, R., Pai a, R. P., and Panda, R. (2024a). A Compa ison S udy o Deep Lea n- ing Me hodologies o Music Emo ion Recogni ion. Sen- so s, 24(7):2201. Lou o, P. L., Redinho, H., San os, R., Malhei o, R., Panda, R., and Pai a, R. P. (2024b). MERGE – A Bimodal Da ase o S a ic Music Emo ion Recogni ion. McFee, B., Ra el, C., Liang, D., Ellis, D., McVica , M., Ba enbe g, E., and Nie o, O. (2015). Lib osa: Audio and Music Signal Analysis in Py hon. In Py hon in Science Con e ence, pages 18–24, Aus in, Texas. O’Malley, T., Bu sz ein, E., Long, J., Cholle , F., Jin, H., In e nizzi, L., e al. (2019). Ke as Tune . h ps:// gi hub.com/ke as- eam/ke as- une . Panda, R., Malhei o, R., and Pai a, R. P. (2020). No el Audio Fea u es o Music Emo ion Recogni ion. IEEE T ansac ions on A ec i e Compu ing, 11(4):614–626. Py o olakis, K., Tzou eli, P., and S amou, G. (2022). Mul i-Modal Song Mood De ec ion wi h Deep Lea ning. Senso s, 22(3):1065. Russell, J. A. (1980). A ci cumplex model o a ec . Jou nal o Pe sonali y and Social Psychology, 39(6):1161–1178. Wa ine , A. B., Kupe man, V., and B ysbae , M. (2013). No ms o alence, a ousal, and dominance o 13,915 English lemmas. Beha io Resea ch Me hods, 45(4):1191–1207.