CNN Performance in Genre Classification on Popular Genres from Amazonia Region: An Analysis in Relation to Human Evaluation and its Implications

CNN performance in genre classification on popular genres from Amazonian region: An analysis in relation to Human evaluation and its implications Claudio Gomes1,2[0000000333554141] and Tsang Ing Ren2[0000000236770264] 1Group Music COmpUter and Technologies Emergent (COUT-e) Computer Science Faculty - Federal University of Amapa - Brazil [email protected] 2Centro de Informática, Universidade Federal de Pernambuco - Brazil [email protected] Abstract. Music is one of the most useful means of communication for social organization, influencing coexistence, customs, lifestyle and tastes. Noticeable in different musical genres, whether for: dancing, celebrating, resting, helping in moments of loneliness, sadness, etc. This work presents adatasetcontainingthemusicalcharacteristicsofthemostpopulargenres from the Amazonia region, from countries of Bolívia, Brazil, Colômbia, Equador, French Guiana, Peru, Dominican Republic, and Venezuela, with 9 popular music genres: andean, brega, carimbó, cúmbia, merengue, pasillo, salsa, vaqueirada, and zouk. This dataset has 22500 samples, 2500 for each genre, segmented in 5 portions of 20 seconds, represented by mel spectrogram. All tracks were acquired by the YouTube or Spotify platform considering viewing rate and sound quality. To evaluate the AMPOP dataset, a custom CNN model was developed for genre classification. Its performance, measured in terms of precision and accuracy, was compared against several state-of-the-art architectures, including MobileNet, ResNet, VGG, EfficientNet, Xception, MobileViT, and MaxViT. Finally, the classification performance of music genre specialists was analyzed based on more than 500 musical evaluations. It was observed that the proposed model significantly outperformed the human experts; however, this difference may be attributed to the inherent subjectivity of human perception as well as the cultural and contextual complexity of the analyzed genres. This highlights the importance of computational approaches to support music genre classification. Keywords: artificial neural networks ·popular music genres ·music information retrieval ·amazonian region. 1Introduction The music industry has been undergoing a profound transformation driven by technological development. Audio streaming service apps like Spotify, SoundCloud, Deezer, among others, have global statistics on listeners’ preferences for Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 209 2C.GomesandT.Ren globally recognized music genres compared to regional or local music genres [16]. These softwares have various tools and strategies that facilitate new music choices or suggestions, driving their continuous use or necessity, such as music genre classifiers, identification of similarities between different music genres, subjectivity models for music genres, and preset equalization [22]. For economic, marketing, content access speed, and other reasons, most of these models mentioned above contain little information about music genres in microregions [24]. As a result, listeners are influenced towards preferences for international genres. In this context, the appreciation of popular music in various countries has gained prominence with the use of artificial intelligence (AI) technologies, which drive cultural preservation, reinvention, and dissemination. In China, for example, AI is used to digitize and reinterpret folk songs, creating modern arrangements that bring new generations closer to ancient traditions [5]. In Nigeria, Machine Learning (ML) is used to classify Nigerian music genres based on musical tracks segments, enhancing the export of Afrobeats and other local styles by analyzing data to reach international audiences and diversify its influence [9]. The scarcity of categorized data in musical databases from diverse regions remains a persistent challenge; however, it also reflects and supports cultural diversity and the interplay between local and global musical influences. Moreover, it plays a crucial role in documenting and preserving the rich cultural heritage of these regions for future generations, while enabling the analysis of evolving sonic trends over time. Amazonia musical genres characterized by their distinctive beats and vibrant rhythms have influenced popular music worldwide, thereby enriching contemporary musical landscapes. This cultural interplay highlights both the universality and the global significance of amazonian contributions to music. However, the classification of lesser known musical genres presents several challenges, including the scarcity of labeled datasets, the structural heterogeneity of sound profiles, and the inherent complexity involved in capturing their unique musical attributes. This work presents the AMPOP a new dataset from popular music genres from Amazonia region. The AMPOP was created containing popular music genres from Amazonia region from the countries: Bolivia, Brazil, Colombia, Ecuador, French Guiana, Peru, Dominican Republic, and Venezuela, considering the main popular musical genres: andean, brega, carimbó, cumbia, merengue, pasillo, salsa, vaqueirada and zouk. After, the dataset was evaluated by AI CNN model to classify these popular music genres. Finally, it was added survey results from musicians experts that classify these genres. 2Relatedworks The development of musical databases has become increasingly crucial in the effort to highlight music from regions traditionally underrepresented in the global music landscape. This initiative seeks not only to diversify the datasets availProc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 210 CNN and Human performance in popular genres from Amazonia 3 able for artificial intelligence research but also to recognize and preserve musical traditions that are frequently marginalized by dominant industry forces. As AI tools continue to advance, there is growing awareness of the need to mitigate biases in recommendation systems and genre classification models, which often favor western centric styles and cultural norms. Expanding collections to include music from underrepresented regions enriches the data ecosystem and paves the way for more inclusive models capable of capturing a broader spectrum of sonic characteristics, instrumental textures, musical forms, and sociocultural narratives [28]. Furthermore, this approach carries significant sociocultural implications. By documenting and analyzing music from underrepresented communities, such datasets contribute to the preservation of cultural heritage that is often at risk of erasure. Increasingly, collaborative initiatives involving local researchers, musicians, and technologists are being established that to ensure these musical traditions are captured with authenticity and cultural sensitivity. By highlighting music from geographically and culturally isolated regions, these initiatives contribute to diversity, equitable, and enriched musical landscape within the context of artificial intelligence. The GTZAN dataset is one of the most widely used public datasets for research in automatic music genre recognition. It contains 1,000 audio tracks, each 30 seconds long, balanced across 10 musical genres: blues, classical, country, disco, hip hop, jazz, metal, pop, reggae, and rock [25]. However, the GTZAN presents several important limitations, including duplicate tracks, labeling errors, and distortions in some recordings, which may affect the evaluation and comparison of music classification systems. The LAMA dataset is a collection focused on the classification of global music genres. It comprises audio files in WAV format extracted from YouTube videos, organized into four geographical categories: Latin America, Asia, the Middle Eastern, and Africa [15]. In addition to the original files, the dataset includes unbalanced 1-minute audio samples, as well as training data based on MelFrequency Cepstral Coefficients (MFCCs) and graphical representations such as FFTs, STFTs, and waveforms. LYRA is an audio-visual dataset dedicated to traditional and popular Greek music. It includes metadata annotations such as genre, instrumentation, title, source, region, and geographical coordinates. This study examined the relationships between regions by analyzing musical representations in conjunction with genres and instruments [21]. PMG-Data is a Persian music genre classification dataset built using CNNs [7]. The dataset contains 500 music tracks from five stratified genres. PMGData includes metadata such as genre, title, and spectrogram. The tracks were segmented into six parts of 8 seconds each, totaling 3,000 samples, in an attempt to improve accuracy. However, it achieved only 65% accuracy, as the 8-second segments did not provide sufficient information for proper genre identification, limiting the model’s learning capacity. Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 211 4C.GomesandT.Ren Model Size Genres Time Input *BL Chen[5] 5000 5 Full MFCC Yes GTZAN 1000 10 30s Audio Yes LAMA 2267 4 60s Audio/AF No LYRA 1570 5/32 10s Mel No PMG-DATA 3000 5 8s Audio Yes AMPOP 22500 9 20s Mel Yes Table 1: Related datasets.*BL: Balanced;*AF: Acoustic features vector means; *NI: Not Informed. Genres Raw Time (h) Andean Youtube/Spotify 20.61 / 10.65 Brega Youtube/Spotify 20.75 / 08.07 Carimbó Youtube/Spotify 09.20 / 20.34 Cumbia Youtube/Spotify 25.60 / 05.89 Merengue Youtube/Spotify 22.25 / 14.79 Pasillo Youtube/Spotify 13.08 / 14.53 Salsa Youtube/Spotify 24.80 / 17.61 Vaqueirada Youtube/Spotify 14.62 / 19.92 Zouk Youtube/Spotify 18.85 / 16.47 Total 169,76 / 128,27 298,03 Table 2: Raw hours from audio tracks on Ampop dataset. Table 1 presents a summary of related works together with the AMPOP dataset. It can be observed that AMPOP presents a larger sampling per genre compared to other datasets, representing greater variety in information with the perspective of improving robustness in training and accuracy in computational models. Furthermore, it is noted that some datasets are not balanced, which can create a negative bias for classification generalization and accuracy. 3AMPOP This case study presents a Music Information Retrieval (MIR) dataset based on audio segments from popular genres native to the Amazonian region, convering countries such as Bolivia, Brazil, Chile, Colombia, Ecuador, Peru, French Guiana, the Dominican Republic, and Venezuela, as shown in Figure 1. The selected genres include andean, brega, carimbó, cumbia, merengue, pasillo, salsa, vaqueirada, and zouk, each representing culturally significant musical expressions within their respective regions. The next subsection presents these genres and their particularities. Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 212 CNN and Human performance in popular genres from Amazonia 5 Fig.1: Case Study of Popular Music Genres in the Amazonian Region 3.1 Particularities of popular music genres Andean music represents a traditional genre of the indigenous peoples of the Andes mountain range, characterized by the use of instruments such as the charango, quena, and zampona, which produce introspective melodies deeply intertwined with cosmological and natural symbolism [4]. Rooted in ancestral cultures of countries such as Peru, Bolivia, and Ecuador, Andean music traces its origins to pre-Columbian traditions and continues to serve as a source of inspiration for contemporary artists [26]. The brega music genre is characterized by romantic lyrics and engaging melodies, often combining elements of tropical rhythms with electric guitars and keyboards with indigenous influences. Popular in the northeast and, particularly, in northern Brazil, it is a symbol of local culture and a style that has entertained parties and generations [1]. Carimbó is a traditional musical genre from the Amazonia region, in the state of Pará, characterized by its vibrant percussion, wooden drums, and circular dances with sensual movements. Originating from indigenous, African, and European influences, carimbó was recognized as Intangible Cultural Heritage of Brazil by IPHAN in 2014 [3], [12]. Originally a coastal cultural expression, cumbia spread throughout Latin America, gaining local variations and becoming a symbol of popular music on the continent [14]. Cumbia is a traditional musical genre and dance originating from Colombia, combining African, indigenous, and Spanish influences, marked by vibrant and percussive rhythms [23]. Merengue is a musical genre from the Dominican Republic, with fast rhythms, use of tambora, güira, and accordion, reflecting African, European, and Caribbean influences [13]. It has big variations in Venezuela and Mexico, becoming a symbol that has gained worldwide popularity due to its contagious energy [11]. Pasillo is a traditional musical genre originated in the Andean region, particularly prominent in Ecuador and Colombia. Characterized by smooth, and Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 213 6C.GomesandT.Ren nostalgic melodies, it reflects a syncretic musical form that integrates European influences most notably the waltz with local cultural and stylistic elements. Usually performed with guitar, mandolin, and other acoustic instruments, it is considered a symbol of Andean cultural identity [2] [27]. Salsa is a musical genre and dance with Caribbean roots, especially Cuban and Puerto Rican, that combines Afro-Cuban rhythms, jazz, and Latin music in energetic arrangements. Emerging in the 1960s in New York, it became a cultural movement and symbol of identity for the Latin diaspora [17]. Vaqueirada from northern Brazil is a cultural expression that adapts elements of northeastern tradition to Amazonian influences, celebrating rural life and the universe of cowboys in events that include music, dance, and competitions [6]. The musical genre Zouk originated in the Caribbean islands, particularly in Guadeloupe and Martinique, which are French territories located near French Guiana. Emerging in the 1980s, zouk represents a fusion of traditional Caribbean rhythms, such as Haitian compas, with influences from Afro-Caribbean calypso. The term zouk means “party” in Antillean Creole and is characterized by smooth, engaging beats and a fluid, sensual dance style that reflects Caribbean culture and celebration [19]. Although zouk originated in the French Antilles, its influence extended to nearby regions such as French Guiana and eventually reached Brazil, particularly the north and northeast regions, where it was adapted and mixed with local rhythms [20]. The Amazonia region, known for its great geographical diversity spanning several countries, rich in its fauna, flora, and social ecosystems, also presents music as a great source of wealth and complexity. Whether due to territorial extension or dialects, Amazonian musicality presents totally different proposals, tonalities, and rhythms, with these selected musical genres being icons in local musical production, influencing recent popular genres [10]. Music does not stop with time. These genres show great variation and musical influences, preserving their uniqueness while exploring new musical possibilities. For example, brega incorporates entirely electronic instruments mixed with other musical genres at a faster pace. Both brega and cumbia embrace musical variations; cumbia blends with carimbó, salsa, and electronic music. Carimbó adapts to include traditional instruments like drums and electric guitars [14]. Salsa has been merging with merengue since the 1990s, and Andean pop music blends with cumbia. Vaqueirada incorporates rhythms from carimbó and andean music. In this way, it is evident that this work highlights how the diversity of the Amazon region is linked not only to biodiversity but also to musical and bodily language. Although separated by vast forests, winding rivers, or distinct origins, these genres share common roots with significant local cultural adaptations. 3.2 Dataset construction The musical dataset consists basically of a computational representation (e.g., mfcc, mel, cqt, tempogram) derived from the content of audio/visual tracks/clips (audio, midi, video) stored in a structured dataset [18]. Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 214 CNN and Human performance in popular genres from Amazonia 7 Fig.2: Systematic Approach to Dataset Creation and Evaluation In most cases, a musical audio track exhibits significant structural variations throughout its duration, including elements such as the introduction, chorus, bridge, and ending. As a result, the temporal dynamics of the signal can influence the measurement of various audio features, such as the spectrogram, tonality, tempo (BPM), and sound intensity. Consequently, it is advisable to analyze the track in shorter, segmented intervals rather than relying on a single, full-track evaluation to ensure more accurate and representative feature extraction [8]. Thus, as shown in figure 2, each musical track was segmented into 5 nonoverlapping 20 second segments. Preliminary experiments were conducted to evaluate the trade-offbetween total segment duration and classification accuracy for dataset construction. In this study, all reported results are based exclusively on these 20 second segments for each musical track. In summary, the dataset illustrated in Figure 2 was preprocessed by converting 20 second segments of the musical tracks into mel spectrogram representations. The audio tracks were sourced from the streaming platforms YouTube and Spotify. The libraries PyTube and Spotipy were employed to download the selected music tracks, followed by conversion to WAV mono format at 22,050 Hz using the ffmpeg tool. Subsequently, the audio segments were preprocessed through normalization to a range of 0 to 1024. The LibRosa3library was used to calculate the mel spectrogram with 224 bands by 672 samples, with fft, window, and hop width of 2048, 1728, 656, respectively. The Mel spectrogram was selected as a representation because it captures all the spectral information that has great potential to identify musical variations in classifications with a lower processing cost [28]. Furthermore, the use of Mel spectrogram is appropriate with the use of CNN’s and ViT’s models that can extract features as spectrogram image references. Musical tracks exceeding 5,000 views and exhibiting acceptable audio quality were selected, while recordings containing noise, silence, or live performances were excluded. The dataset encompasses over 423 artists, with recordings spanning from 1947 to the present, thereby reflecting both the historical continuity and variability of the selected genres. All tracks were assessed and validated by experts to ensure accurate classification within each musical category. The primary objective of this study was to classify the selected musical genres, with an emphasis on achieving maximum accuracy across computational models. To this end, the dataset was partitioned into training, validation, and test subsets, comprising 70%, 15%, and 15% of the data, respectively. Given that 3LibRosa: https://librosa.org/doc/latest/index.html Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 215 8C.GomesandT.Ren Fig.3: AMPOP CNN architecture. *F: Filters; *K: Kernel; *S: Size. the dataset comprises five 20 second segments from each musical track, as illustrated in Figure 2, all segments from a single track were assigned to the same subset training, validation, or test during random selection. This dataset contains mel spectrograms derived from 20 second segments of publicly available music recordings sourced from YouTube and Spotify. The data derived are shared solely for academic research purposes under fair use and with appropriate citation. The dataset is provided exclusively for academic and non-commercial use, with the aim of promoting research in music information retrieval and computational musicology. No content is included that allows for the reconstruction of the original recordings. The dataset is available via a link4. 3.3 AMPOP CNN genre classifier Table 3 reports the validation and test classification outcomes for the dataset, obtained using the optimal optimizer across the AMPOP, EfficientNet, MobileNet, ResNet, VGG16, Xception, MobileViT, and MaxViT architectures. Additionally, the corresponding accuracy, precision, and the hyperparameters selected during the training and validation phases are reported. Model ATE(s) ICR(MB) Valid(%) Acc |Prec Test(%) Acc |Prec Opt EfficientNet 355 7988.38 81.34|81.42 77.88|78.59 adam MobileNet 218 6392.76 85.08|85.13 72.15|72.25 adamax ResNet 679 46118.58 84.28|84.19 79.38|80.27 adam VGG 1355 28804.76 91.36|91.48 83.42|83.78 adam Xception 679 40895.71 94.89|95.02 86.95|87.45 rmsprop MobileViT 459 34212.08 12.50|12.43 12.50|12.43 adam MaxViT 3577 26787.42 13.92|13.88 15.08|06.12 rmsprop AMPOP 221 219.51 96.77|97.48 87.21|87.43 lion Table 3: Evaluation results on validation and test datasets. Acc: Accuracy; Prec: Precision; *ATE: Average Time by Epoch; *ICR: Initial Computational Resources. 4AMPOP dataset: https://github.com/claudiorogerio/AMPOP Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 216 CNN and Human performance in popular genres from Amazonia 9 (a) Results by segments (b) Accuracy(A) and Precision(P) (c) Predictors by genres (d) Ground truth by genres Fig.4: Test Results with the AMPOP Model Systematic analyses were conducted on variations in hyperparameters and optimizers. While the model architectures remained constant across configurations, the resulting accuracy and precision exhibited substantial variation depending on the optimizer applied. Based on the results presented in Table 3, all corresponding hyperparameters were held constant to conduct a final evaluation on the test dataset and assess the resulting accuracy and precision. Astrategywasfocusedonmaximizingaccuracyandprecision,forallmodels. A batch size was of 256, with optimizer variations as Adam, Adamax, AdamW, LION, Nadam, RMSProp, and SGD. Training was configured with a maximum of 200 epochs, with early stopping and an initial learning rate of 0.001, subject to an exponential decay of 80% every 40 epochs. The AMPOP model was observed to achieve the highest accuracy and precision on the test set, followed by Xception, which yielded the second best performance. However, it is important to highlight that AMPOP has demonstrated a lower Average Time per Epoch (ATE), measured in seconds, and reduced Initial Computational Resource consumption (ICR), measured in MB in training set, eliminating the need for high-performance computing infrastructure. Consequently, AMPOP emerges as the most suitable architecture for the classification of popular music genres from the Amazonian region, a conclusion supported not only by its superior accuracy and precision but also by its lower computational Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 217