Distorted Realities: Classifying Extreme Vocals Between Harmony and Noise – A Machine–Human Evaluation of Vocal Confusion Patterns

3.1 Existing taxonomies and classification challenges

Currently, most classification systems in the extreme metal field are based on qualitative, practice‑based observations rather than being grounded in systematically quantified acoustic parameters (Tailleur et al., 2024).³ Sadolin (2012: 177–202) classified some lighter vocal techniques found in extreme metal, such as Distortion, Rattle, Growl and Grunt. Nieto (2008, 2013) introduced the concept of ‘Vocal Effects’ or ‘Extreme Vocal Effects’, categorising extreme vocals into three groups: (1) Growl, common in death metal and other extreme music genres, is highly noisy, with fundamental frequencies rarely perceived, and it involves significant spectral variation; (2) Fry Scream, induced by irregularly spaced glottal pulses during inhalation or exhalation, is commonly known as Pig Squeal and is similar to Growl but with a brighter spectrum; and (3) Roughness, a category that includes techniques that add subtle variations to the vocal tract, produces a harmonically richer spectrum and is more commonly found in hard rock.

Wallmark (2018) likewise summarises distorted vocals in death metal under the umbrella term ‘death vocals’, which includes growls, grunts, pig squeals, screams and deep gutturals. He describes these as predominantly low‑register singing styles characterised by a thick, raspy and heavily distorted vocal quality. He also notes that singers conceptualise their tonal options in broad register categories – namely highs, mids and lows (see Wallmark, 2018: 72). In the MVD (Kalbag and Lerch, 2022), the authors proposed that most screams in modern metal are variations of Fry Scream. Their dataset mainly focuses on the classification of Fry Scream, dividing it into three perceptual categories: high, mid and low. In scene‑informed discourse, Kennedy (2018) divides distorted vocals into three broad categories, growls, screams and roars, and relates these categories to coarse frequency tendencies and genre associations. The author also notes that boundaries between growls and roars can be malleable and often indistinct, with roaring occupying an intermediate region between low growls and high screams or shrieks (Kennedy, 2018).

Tailleur et al. (2024) introduced a more comprehensive classification that not only covers vocal techniques but also emphasises their association with specific extreme subgenres. Their classification is divided into: (1) Vocal Techniques, defined by the ability to replace a clean voice, articulate understandable lyrics and adjust pitch within a musical range (e.g. Black Shriek, Death Growl, Hardcore Scream and Grind Inhale), and (2) Vocal Effects, defined by the incomprehensibility of lyrics and their typical use as embellishments in performances (e.g. Pig Squeal, Deep Gutturals and Tunnel Throat). In this classification, vocal techniques are primarily methods of sound production that can replace clean vocals while maintaining pitch control and lyric articulation. In contrast, vocal effects are embellishments that often obscure lyrics while serving as textural or expressive devices.

In addition to classification methods specific to extreme metal music and related genres, there are other broader classification frameworks for extreme vocals, which may be applicable to this research. For example, classification can be based on anatomical structures and phonatory mechanisms, such as the vibratory phonation of the ventricular folds, vocal folds and aryepiglottic folds mentioned earlier, as well as the taxonomy of Supraglottic Structure Vibrations; see Aaen et al. (2025) for discussion. A classification based on resonating cavities and airflow physiology can also be considered: Extreme vocals can be produced using either exhalation or inhalation, resulting in two distinct styles (Kalbag and Lerch, 2022; Smialek et al., 2012).

Finally, during vocal performance, singers can also modify their phonation either by lowering or raising the jaw, which affects the perceived pitch and timbre of extreme vocals (Smialek et al., 2012) and can be used as a classification criterion.

3.2 Practice‑Based taxonomies: Interview data

In practical contexts, classification among practitioners is often experience‑based. To capture practice‑grounded perspectives, we conducted exploratory interviews with 11 experienced extreme metal vocalists (see Appendix 2 for transcripts). Interviews were conducted remotely in an asynchronous format via online messaging. Participants responded in writing and, when preferred, with short voice messages, allowing them additional time to reflect on the questions and formulate their answers. The first interview used open questions about possible classification principles and phonation strategies. Based on this first interview, in which the vocalist proposed a high‑, mid‑ and low‑frequency–based grouping, we later asked all vocalists how they viewed this classification. Even when one or two participants stated upfront that they themselves already classify by frequency, we still posed the same question to maintain a consistent interview protocol and enable comparable responses across participants. As noted by Fan, the vocalist of Frozen Moon and one of the managers of Pest Productions, the largest black metal label in Asia, during a Personal Interview (see Appendix 2A):

Dyingflames, the owner of Diensysian Records and a musician in Holokastrial and R.N.V., expressed a similar opinion:

An anonymous black metal singer agreed with their perspective:

Another vocalist, Jovi, who has extensive experience in cover performances, proposed a slightly different classification system in a personal interview (see Appendix 2E), suggesting that extreme vocals could be categorised into four groups: Recitation, Clean Singing, Melodic Singing and Distortion. Three metal vocalists (20+ years experience) supported frequency‑based classification but noted that fixed frequency boundaries are hard to define. Additional practitioners expressed a more style‑oriented perspective. GUT from Huangquan Records and the death metal band Mvltifission argued that classification should primarily follow stylistic context, since a simple ‘high, mid, low’ scheme is too general and each style itself can contain high, mid and low vocal approaches (see Appendix 2F). DomAs666 from Naked Whipper agreed with this position and considered a style‑based classification more flexible for both musical and instrumental practice, allowing the voice to produce a wider range of sonic effects (see Appendix 2I). In his view, the traditional ‘high, mid, low’ framework is workable but overly rigid for practical artistic use. However, other interviewees held a different opinion and regarded the ‘high, mid, low’ system as more practical overall, arguing that, when the underlying production principles are similar, further subdivision is not necessary.

As Fan noted (see Appendix 2A), there are many vocal techniques, and a fully exhaustive organisation is difficult to achieve. This difficulty is compounded by the fact that, as Brackett (2016) points out in his discussion of the concept of genre, genre boundaries are themselves porous and blurred, and overlap and intersection constantly occur between closely related subgenres (see Chapter 1 in Brackett, 2016 for discussion). A similar pattern can be observed in extreme music, especially in the late 1980s and early 1990s, when the stylistic practices of many bands did not strictly conform to a single genre framework but instead simultaneously displayed features of multiple extreme music styles. For example, in the works of the Brazilian bands Vulcano and Sarcófago, traits associated with thrash metal, death metal and black metal are often intertwined. Similarly, the boundaries between grindcore, punk and metal are far from clear. In the case of Napalm Death, one can hear elements of grindcore, hardcore punk, death metal and thrash metal across different periods or releases; likewise, the early form of grindcore displayed by Extreme Noise Terror also reflects, to a considerable extent, the continuity and hybridity between hardcore punk and metal.

In such contexts of genre overlap or hybridity, the vocal styles used by bands may also vary considerably, depending to a large extent on the particular vocalist’s phonatory habits and technical approach; at the same time, when the other members and the overall compositional orientation remain relatively unchanged, the musical style of the band may not necessarily undergo a fundamental shift as a result.⁴

This suggests that there is no simple one‑to‑one correspondence between genre labels and vocal labels; a single genre may accommodate markedly different vocal realisations, while similar vocal sound results may also occur in different stylistic contexts. Although GUT and DomAs666 suggested in our interviews that classification could follow stylistic context, this is not always sufficient. Kennedy (2018) notes that genre framing can itself rely on contrasts between clean singing and distorted vocals – for example, in progressive death metal and related contexts. For this reason, we adopt the high‑, mid‑ and low‑frequency–based grouping as a first step.

Based on insights from these interviews and the recommendations of the vocalists, our research group⁵ analysed 117 audio excerpts of punk, metal and hardcore vocals, provided as part of a seminar in the musicology program at the University of Cologne.⁶ The excerpts have a duration of 6.2 s on average (range 0.6–22.8 s, N = 117). The simple frequency‑based classification (high, mid, low) was operationalised by assigning excerpts to three frequency groups via Statistical Package for the Social Sciences (SPSS) automatic binning of maximum amplitude frequency (see Section 4.1, Footnote 7). Though seemingly intuitive or subjective, the consistent recurrence of the simple frequency‑based classification across interviews indicates perceptual coherence. Rather than treating it as a fixed taxonomy, we use it as an initial sanity check to explore whether such performer‑informed distinctions align with measurable acoustic features. This perspective allows us to contribute to ongoing efforts to bridge the gap between practice‑based knowledge and computational modelling in extreme vocal analysis. Although vocal techniques and effects are often discussed in terms of acoustic features, some practitioners distinguish them primarily by articulatory technique, regardless of the resulting sound.

Interviews revealed that, while listeners use frequency cues to distinguish vocal types, techniques like ‘Black Shriek’ and ‘Hardcore Scream’ differ in vocal production. These differences are genre‑specific and involve distinct physiological mechanisms (see Section 4 ‘Methodology’). While listeners familiar with the respective genres (e.g. black metal or hardcore) often identify these techniques by ear, performers themselves emphasise that they rely on different vocal‑production strategies. Vocalists learn them through embodied experimentation and recognise them within shared stylistic frameworks. Some techniques (e.g. growl and pig squeal) are documented, while others evolve informally as genre‑specific or hybrid forms. Though rarely formalised in scholarly literature, performers consistently acknowledge them as distinct and reproducible within their respective traditions. As such, performer intuition and stylistic convention are indispensable in understanding the functional boundaries of extreme vocal techniques.

The interview‑derived taxonomy offered insight into how performers intuitively group vocal styles. However, because such groupings are inherently dependent on the vocal range of a given singer and lack cross‑performer standardisation, they were considered unsuitable as ground‑truth labels for computational modelling in the later phase of this study. Instead, we treated them as a perceptually grounded probe that informs the acoustic validation and serves as a reference framework for the ML analyses presented in Section 4.

3.3 Singing techniques

As discussed in the previous section, current extreme vocal‑classification frameworks are largely practice‑based and historically developed, which leads to inconsistencies across taxonomies. To empirically evaluate these categories, we conducted a set of perception and ML experiments using the EMVD by Tailleur et al. (2024). The dataset provides operational class labels for supervised training and evaluation. In the following subsections, we report the human‑perception experiment and the subsequent ML analyses (see Section 4), including the main classification model as well as auxiliary analyses, such as feature‑group importance and ablation tests. A description of the EMVD is provided in Section 4.2.

To better understand the data evaluated and the results achieved, we provide a brief introduction to the four vocal techniques used in this study. Figure 1 displays log‑mel spectrograms generated using the Python library Librosa by McFee et al. (2015). Log‑mel spectrograms were chosen because they align better with human auditory perception, i.e. processing frequency distributions similarly to how humans naturally perceive sound (Jothimani and Premalatha, 2022; Li et al., 2024). The spectrograms show audio from the EMVD recorded by singer ID 10 (male) performing ‘Mid_a’, i.e. the vowel [a] performed in a mid‑frequency range, using each of the four vocal techniques. In the EMVD dataset, Black Shriek, Death Growl and Hardcore Scream are categorised as Vocal Techniques. These are alternative vocalisations to Clean Voice, capable of conveying intelligible lyrics while allowing for stylistic variations. Supraglottic constriction plays a significant role in extreme vocalisation techniques (Eckers et al., 2009; Erbe, 2014; Sakakibara et al., 2004), making it a key distinguishing feature between clean voice and other vocal styles. As clearly visible in the spectrograms (see Figure 1), clean voice can readily be distinguished from other techniques, exhibiting distinct and stable harmonic lines, particularly concentrated in the lower frequency range. For a reproducible comparison across techniques, the white box marks the main band, defined as the 20th to 90th percentile frequency interval of the time averaged mel band energy over the displayed excerpt.

Figure 1

Black Shriek demonstrates comparatively stronger energy in the upper frequency region than the other techniques, with a dense and sustained appearance in the spectrogram. From an auditory perspective, Black Shriek (from now on also referred to as Shriek for simplicity) resembles a harsh, high‑frequency noise like vocal quality with relatively limited tonal salience. This type of sound is widely used in black metal, including canonical bands such as Darkthrone and Emperor, and can also be heard in crossover acts such as Violent Magic Orchestra. In the EMVD metadata, it is classified as an exhaled technique and is generally described as higher than Death Growl (Tailleur et al., 2024). In a personal interview (Appendix 2C), Dyingflames indicated that Shriek and Scream share similar vocal‑production techniques: ‘High[‑]pitched extreme vocals are mostly similar in production. Shriek has a more pronounced grainy texture, and during production, just like with Scream, the mouth is opened in a smile‑like shape, and the throat is compressed’.

Death Growl, on the other hand, is perceived as a low‑frequency sound with minimal human vocal characteristics, resembling animalistic growling. Growl is common in death metal and can be heard in bands such as Deicide, Morbid Angel, Pestilence, Grave, Death and Suffocation. Its spectrogram shows energy predominantly concentrated in the low‑ to low‑mid‑frequency range, appearing dense but somewhat blurred and lacking sharp, high‑frequency features. An endoscopic study (Eckers et al., 2009) revealed two primary ways singers produce this technique: (1) vibration of the ventricular folds accompanied by an anteroposterior constriction of the supraglottic tract and (2) vibration of the aryepiglottic folds, which produces a sharper, squealing sound similar to a Pig Squeal. Erbe (2014) shows that some singers performing Growl do not use vocal fold vibration; instead, they maintain their vocal folds in a rigid state and rely on vibrations from other anatomical structures. In some cases, the epiglottis strikes and makes contact with the arytenoid cartilages. Additionally, mucus or saliva may vibrate during performance, contributing to the roughness of the vocal timbre while simultaneously providing a protective effect for the vocal folds. Wallmark further notes that, during phonation, singers should keep the throat well hydrated, avoid excessive throat tension and sing with diaphragmatic support (see Wallmark, 2018: 75). Dyingflames, known for his proficiency in Growl, explained: ‘Growl is different from the other techniques. The mouth forms an “O” shape, and the vocal folds lower – something like that’ (see Appendix 2C).

Hardcore Scream occupies mid‑ to high‑frequency ranges and retains partial human voice characteristics. The spectrogram clearly shows layered frequency bands from low frequency (approximately 128 Hz) to mid‑frequency (approximately 2048 Hz), reflecting distinct vocal resonance layers. Interviewee B described this vocal technique as ‘the vocal cords coming together, with airflow pushed upward, forcing a little sound through tiny gaps’ (see Appendix 2D), which illustrates the characteristics observed in the spectrogram from an audible point of view. A similar vocal quality can be heard in the music of The Blood Brothers and Slant. It is also present in the vocals of Fuming Mouth, where Mark Whelan shifts between Scream, Growl and Clear Voice in his performances. Finally, Dyingflames added, ‘the sound of Hardcore Scream is drier than Black Shriek’, although he noted that the production technique is essentially the same (see Appendix 2C).

However, the term Hardcore Scream remains open to discussion, as it does not necessarily denote a homogeneous category in scene discourse. Kennedy (2018) describes early hardcore vocals as closer to an overdriven shout and discusses practices such as group vocals, illustrating that hardcore vocal delivery is not reducible to a single technique and that genre context shapes how vocal descriptors are used across hardcore and metal‑related styles. In Kennedy’s account, screaming is often associated with grindcore and black metal contexts, yet our material suggests that label usage and acoustic tendencies do not map one to one across corpora and communities. In particular, Figure 1 and several interview comparisons indicate that the present dataset categories capture distinctions that are context‑dependent and may differ from genre level descriptors used elsewhere (see Appendices 2C, 2D, 2F and 2I). Therefore, we retain the EMVD category names as operational labels for the current study and follow the dataset annotation by treating Scream as Hardcore Scream in the experimental sections.

4.1 Taxonomy validation

As an initial step, we preprocessed and analysed the 117 audio excerpts provided in the University course to evaluate whether the frequency‑related taxonomy introduced in Section 3.1 and derived from practitioner’s interviews corresponded to measurable acoustic differences.⁷ Each clip was transformed into the frequency domain using a discrete Fourier‑transform implemented through SciPy FFT (fast Fourier transform); see Virtanen et al. (2020). The transform was applied once to the full‑length waveform (i.e. no short‑time framing), yielding a single‑sided magnitude spectrum in Hz.⁸ To assess the validity of the interview‑derived categorisation, the samples were divided into three groups, i.e. high‑, mid‑ and low‑frequency, based on their maximum amplitude frequency (max_frequency), defined as the frequency location of the maximum spectral magnitude.

We conducted a one‑way analysis of variance (ANOVA) to compare the group means. Because the assumption of homogeneity of variances was violated (Levene’s test ($F(2,92)=4.42$, $p=0.015$)), we used Welch’s ANOVA, which revealed a significant group effect ($F(2,58.41)=37.32$, $p<0.001$) with a large effect size (${\omega }^2=0.47$, 95% CI [0.31, 0.57]). Post‑hoc pairwise comparisons, evaluated using Games–Howell tests, revealed that all pairs were significant ($p<0.001$): the high‑frequency group differed from the mid‑frequency group by about 291 Hz and from the low‑frequency group by about 535 Hz; the mid‑frequency group also differed from the low‑frequency group by about 245 Hz. These results support the validity of our taxonomy‑based frequency groups in terms of maximum amplitude frequency. To complement the analysis based on the maximum‑amplitude frequency, we repeated the validation using the spectral centroid (Hz).⁹ Since the assumption of homogeneity of variance was again violated (Levene’s test: $F(2,114)=12.54$, $p<0.001$), Welch’s ANOVA was applied. The ANOVA results (Welch’s $F(2,65.0)=141.00$, $p<0.001$, ${\omega }^2=0.75$, 95% CI $[0.67,0.80]$) indicate a very strong group effect. Games–Howell post‑hoc tests further revealed significant pairwise differences between all groups (all $p<0.001$): the high group exceeded the mid group by approximately $858$ Hz and the low group by approximately $1569$ Hz; the mid group exceeded the low group by approximately $711$ Hz. These results closely mirror the max_frequency analysis, which supports the validity of our frequency‑based taxonomy.

The bar chart in Figure 2 illustrates the average max_frequency values for each group. The error bars indicate a relatively small margin of error, suggesting that using frequency‑based grouping is a feasible approach for our fairly stable set of 117 excerpts.

Figure 2

While perceptually coherent, the frequency‑based taxonomy varies across singers, as corresponding frequency ranges differ individually. This makes it unsuitable as a standard ground truth for supervised learning. We therefore used it solely as a reference for interpretation and performer insight. That is, the frequency‑based taxonomy is discussed qualitatively, whereas the supervised models in the following sections use the EMVD labels as operational classes for supervised training and evaluation.

4.2 Dataset and evaluation metrics

In this work, we use the EMVD by Tailleur et al. (2024),¹⁰ consisting of 1068 audio instances (approximately 100 min of recordings) of eight different vocal techniques performed by 27 singers (including four women). Due to uneven distribution of samples across categories, our experiments focused exclusively on the four vocal techniques with at least 100 instances per class – Clear Voice, Black Shriek, Death Growl and Hardcore Scream – for a total of 946 audio instances (for its distribution, see Table 1). In the EMVD dataset, each audio recording is rated on a scale from 0 to 2 points, where 0 points indicates insufficient representation of the intended vocal technique, i.e. unsuitable for deep learning (Tailleur et al., 2024); 1 point indicates partial representation with moderate quality; and 2 points indicates excellent representation, closely matching the expected technique.

To assess human perception and ML testing, the same audio samples were considered (from now on, referred to as the ‘test set’). Only samples awarded 2 points were selected to ensure reliability. In addition, samples from the ‘lyric’ category of the dataset were chosen for this set: Each audio sample in this category is roughly 15 s long, with lyrics chosen by the singers themselves, consistently used across all vocal techniques. The final test set included 71 samples by 23 singers. Four additional samples by different singers were used as examples during the human‑perception experiment but were not included in the test set.

To evaluate both human perception and ML classification, we use unweighted average recall (UAR), precision and recall. UAR, also known as balanced accuracy, is the average recall across all categories and is more suitable than traditional accuracy for handling imbalanced class distributions (Bekkar et al., 2013). We also present confusion matrices to illustrate confusions among categories.

4.3 Listening test

We targeted a group of participants with prior auditory experience related to extreme vocal techniques. A total of 158 participants aged from 18 to 56 years (M = 29.48, SD = 8.05), including 23 females, 123 males, five non‑binary/diverse individuals, four individuals who preferred not to disclose their gender, and three individuals who left the field blank, were included in the analysis (see Appendix 3 for the participant flow diagram). They were mainly recruited through the following channels: (1) niche music enthusiast communities, (2) metal club membership and (3) public online announcements. The procedures used adhere to the tenets of the Declaration of Helsinki.¹¹

After signing informed consent forms guaranteeing the use of their anonymous responses for research, participants first listened to demonstration audio samples from the four vocal categories (Clear Voice, Black Shriek, Hardcore Scream and Death Growl). These four demonstration samples served as canonical examples to familiarise participants with the task and were not included in the evaluation set. They were selected from the EMVD dataset as high‑quality, prototypical instances, according to the assigned grade given in the dataset, i.e. rating = 2 points (Tailleur et al., 2024). To minimise potential speaker‑identity cues, we selected the four demo exemplars manually under three constraints: each exemplar had a rating of 2 points according to the EMVD quality grade, the four exemplars came from four different vocalists and we avoided using the same vocalist for the demo exemplar and evaluation items of the same class wherever possible.¹² Participants were asked the following question for each audio clip: ‘Which vocal technique do you think is used in this clip?’ Then, they annotated 20 randomly‑assigned audio excerpts in a force‑choice task where they could select only one category. To avoid the influence of genre associations on the participants, we removed genre indicators from the voice‑type labels, as shown in Figure 3.

Figure 3

We conducted the experiment online using the SoSciSurvey platform. A total of 71 audio excerpts were evaluated, including 11 Clear Voice, 20 Shriek, 21 Growl and 19 Scream samples. Since previous experiments demonstrated listeners’ ability to distinguish between clean voice and distorted voice quite effectively (Tailleur et al., 2024), we included fewer clean voice samples. In each round, to preserve accuracy and avoid overloading participants, only 20 randomly assigned excerpts were assessed by each individual. Stimuli were presented using a within‑subjects randomisation design implemented in SoSciSurvey such that each participant evaluated 20 randomly selected audio clips drawn from the same stimulus pool. Due to the random‑assignment procedure implemented in SoSciSurvey, individual audio clips were not presented an equal number of times; across the 71 clips, the number of annotations per item ranged from 49 to 59 (M = 53.64). Each sample was annotated by at least 48 participants. Upon completing the listening test, participants filled out demographic information, including their gender, age, preferred music genre(s), type of auditory experience with extreme vocal techniques and the number of years they had been exposed to such music. Finally, participants could share any additional thoughts or request feedback on their responses through an open‑ended field.

As shown in Figure 4, regarding auditory experience, the majority of participants (63.29%) identified themselves as ‘I am a fan/regular listener’. In addition, 28 participants (17.7%) indicated that they ‘perform extreme vocals’, while seven participants reported engaging in music production or recording related to these techniques. Notably, in terms of years of exposure (see Figure 5), 66 participants (41.8%) reported more than 10 years of experience, while only six had less than 1 year of exposure, showing the rich auditory domain experience possessed by the participants.

Figure 4

Figure 5

In terms of music genre preferences, participants showed a wide range of interests. Figure 6 shows the percentage of participants selecting each given genre in the multi‑select genre‑preference question: ‘Which music genres that use extreme vocals do you enjoy listening to?’ (select all that apply). ‘Other’ denotes the share of participants who provided supplementary responses via a free text field. In total, 13.3% selected all five listed music genres, 6.3% not only selected all provided genres but also supplemented their answers by reporting additional preferred music styles under the ‘Other’ category and 25.9% selected two music genres.¹³ Death Metal was the most popular (111 participants), followed by Black Metal (102 participants) and Hardcore (90 participants). The three assessed vocal techniques are often associated with these three genres, which guarantees consistency between the listening setup and the participants’ auditory preferences.

Figure 6

4.4 Machine learning setup

In addition to the main ML classification (i.e. the baseline classification task), we conducted three auxiliary analyses: (i) a feature group importance analysis to quantify the contribution of broad ComParE feature families; (ii) a perception‑oriented human prediction task, in which an additional SVM was trained to predict listeners’ majority decisions using the same feature representation, singer‑independent three‑fold partitioning and a hyperparameter optimisation procedure as in the baseline classification task; and (iii) an ablation analysis probing the role of perturbation and noise‑related cues by comparing a model restricted to Jitter, Shimmer and harmonics‑to‑noise ratio (HNR) features with a complementary model trained on all remaining ComParE features.

4.4.1 Baseline classification task

In audio signal processing, mel frequency cepstral coefficients (MFCCs) and spectrograms have long been standard due to their alignment with human auditory perception (Purwins et al., 2019). Prior work has confirmed their effectiveness for vocal technique recognition (Stadler et al., 2023), including in metal vocal datasets such as MVD (Kalbag and Lerch, 2022) and EMVD (Tailleur et al., 2024), where log‑mel spectrograms were primarily used for convolutional neural network (CNN) input. While binary classification between Clear Voice and extreme vocals yielded high accuracy (93%), multi‑class performance declined significantly (Tailleur et al., 2024), likely due to more subtle spectral distinctions among distorted vocal styles.¹⁴ Additionally, studies have shown that extreme techniques like Death Growl and Scream exhibit higher Jitter and Shimmer, and lower HNR (Kato and Ito, 2013), suggesting that incorporating a broader set of acoustic features could further improve classification performance.

The ComParE feature set (Eyben et al., 2010) is a suitable candidate for our purposes and has shown good performance in similar studies (Parada‑Cabaleiro et al., 2023; Xu et al., 2022). It includes 6373 statistical functionals computed from 65 low‑level descriptors (LLDs) and their delta coefficients (including MFCCs, spectral, prosodic and voice‑quality features). We extracted features using the openSMILE toolkit (Eyben et al., 2010) with default parameters. To reduce feature dimensionality, computational cost and improve model performance (Guyon and Elisseeff, 2003; Kohavi and John, 1997), we use the SelectKBest feature‑selection method from scikit‑learn by Pedregosa et al. (2011), which evaluates the importance of each feature based on the ANOVA F‑test and selects the top k features.

As in similar tasks (Kalbag and Lerch, 2022; Parada‑Cabaleiro et al., 2023; Stadler et al., 2023; Xu et al., 2022), we used a SVM implemented on scikit‑learn by Pedregosa et al. (2011) with a linear kernel. We performed hyperparameter optimisation using a nested grid search within each training fold only, with singer‑wise splits applied inside the training pool, while test singers were fully held out and never used for model selection (Pedregosa et al., 2011); the candidate grids were defined a priori based on preliminary manual tuning (feature numbers [100, 200, 300, 500, 700, 900, 1000]; regularisation parameters [1, 0.1, 0.001, 0.0001, 0.00001, 0.000001]). Additionally, we conducted an exploratory feasibility check with a simple two‑layer CNN using log‑mel spectrogram input and dynamic padding in the nine‑class task (Clear Voice, Black Shriek, Death Growl, Deep Gutturals, Effect, Grind Inhale, Hardcore Scream, Pig Squeal and Tunnel Throat) but observed notably lower performance (micro accuracy 70%; macro average 32%); for parity, we also trained a nine‑class linear SVM with ComParE features and SelectKBest (top 100), achieving a micro accuracy of 77.1% and macro average of 52%. The corresponding nine‑class confusion matrices (CNN and SVM) are reported in Appendix 1, while the main analyses reported in the Results section focus on the reduced four‑class taxonomy.

To avoid the model learning individual singer characteristics and prevent data leakage, we split the data based on Singer ID and adopted a singer‑independent three‑fold evaluation scheme. Test folds were constructed from the human‑evaluation stimulus pool by assigning singers to three non‑overlapping groups using a priority‑queue (heapq) load‑balancing procedure based on the number of human‑evaluated stimuli per singer so that the total number of test stimuli was approximately balanced across folds. We carried out three independent experiments, each time on a different test set containing singers not seen in the training/validation sets (Fold 0 test singers: [2, 9, 10, 11, 12, 19, 23], 24 audio excerpts; Fold 1: [1, 3, 5, 13, 14, 16, 17], 24 audio excerpts; Fold 2: [4, 7, 8, 15, 18, 21, 22, 25], 25 audio excerpts).

Figure 7 depicts the distribution of samples across sets. To ensure comparability with the human‑perception experiment, the accumulative test sets across experiments correspond to the same 71 excerpts assessed in the perception experiment. Groups with more excerpts were merged with those with fewer excerpts to ensure that test excerpts per experiment were within ±2. Table 2 shows the distribution of excerpts across classes in the test set for each experiment. When discussing the ML results (see Section 5.2), we report the mean across the UAR obtained from the three experiments.

Figure 7

Table 2

Distribution of the 71 samples included in the test set across vocal techniques. To guarantee a singer‑independent ML setup, the samples were distributed across three independent experiments with unique singers in the test set.

Technique	Experiment 1	Experiment 2	Experiment 3
Clear Voice	4	4	3
Black Shriek	6	6	8
Death Growl	5	9	7
Hardcore Scream	9	3	7
Total	24	22	25

5.1 Human‑perception experiment

Figure 8 shows the confusion matrix for the human‑perception experiment (UAR of 76.2%). Clear Voice achieved the highest recall (97%) and Growl reached nearly 80%. In contrast, the recall for Shriek and Scream were lower, both below 70% (68.6% and 60.9%, respectively). The precision scores for these two techniques were also relatively low, especially for Shriek (63%), even below Scream (68%), as the former attracted the highest confusion from the other classes (see darker column for Shriek in the confusion matrix reported in Figure 8).¹⁶

Figure 8

All in all, these results indicate that Clear Voice and Growl were more reliably identified by human listeners, while Scream, but in particular Shriek, were more often confused with other categories.

Although Growl and Shriek show some differences in their spectrogram representations (Growl is more concentrated in low frequencies, whereas Shriek is more concentrated in high frequencies), 15.2% of Growl samples were still misclassified as Shriek. The classification performance for Scream was the worst, with the highest confusion with Shriek (23.5%), which may be attributed to the similar frequency characteristics of Scream and Shriek (see spectrograms in Figure 1). One possible explanation for this is that listener expectations and mental‑categorisation strategies influenced their decisions: When exposed to a series of extreme vocal styles, participants may have second‑guessed certain Clean Voice samples, interpreting subtle vocal effort or coloration as indicative of a more distorted technique. This suggests that perception was shaped not only by acoustic features but also by the context in which the samples were evaluated. Moreover, it is possible that some listeners interpreted the Shriek category as a sort of fallback label, assigning uncertain samples there when no clear match was perceived. Such tendencies may indicate that Shriek was perceived as less acoustically distinct compared to other categories.

5.2 Machine learning experiments

5.2.1 Baseline classification task

We conducted three independent ML experiments using a singer‑independent three‑fold group cross‑validation scheme (by singer ID) for optimisation. In each fold, the test singers were fully held out and differed across folds. The reported results are averaged across the three folds. The SVM demonstrates good performance, achieving a UAR ($\mu \pm \sigma$ in %) of $90.68\pm 3.94$. Overall, the pooled test‑set UAR was 90.0% (95% bootstrap CI: [82.9%, 96.3%]; N = 71). The confusion matrix (see Figure 9) illustrates the classification results and confusion patterns. The classification performance across categories was high, with Clear Voice reaching both a precision and recall of 100%. Growl achieved a recall of 85.7% and precision of 89.1%. For Shriek and Scream, even if varying in recall (85.0% and 89.5%, respectively), precision scores were both 85.6%.

Figure 9

In addition, per‑class PR‑AUC values were: Clear Voice $=100.0\%$, Shriek $=91.8\%$, Growl $=87.4\%$ and Scream $=86.6\%$ (macro avg. $=91.7\%$; micro avg. $=91.0\%$). The combination of high precision and recall across all categories suggests that the model is capable of identifying each vocal technique both accurately and reliably, thus outperforming humans not only in overall accuracy but also in consistency across categories.

Clear Voice was classified perfectly, which also confirms previous findings (Tailleur et al., 2024) that distorted and non‑distorted vocals can be effectively recognised by ML models. Shriek achieved a recall of 85%, with 15% of samples misclassified as Scream. This result is consistent with the human‑perception experiment, where 22.4% Shriek samples were misclassified as Scream. The ML model reproduces two salient human‑confusion directions (see Figure 8 versus Figure 9 for comparison). Growl is frequently misclassified as Shriek (ML model 14.3%; human 15.2%). Scream is also misclassified as Growl at a comparable rate (ML model 10.5%; human 11.5%). These alignments suggest that the main errors reflect shared perceptual ambiguities between these technique pairs.

Most notably, human participants frequently confused Scream and Shriek, with Scream being misclassified as Shriek in 23.5% of cases. In contrast, such confusion did not exist in the ML classification results. This may be attributed to the model’s ability to perform fine‑grained feature extraction, using a greater number of acoustic features to detect subtle differences that are often imperceptible to human listeners. Moreover, ML models are not subject to subjective biases and can consistently evaluate clarity and frequency distribution.

5.2.2 Feature group importance analysis

Across folds, the highest mean group weight observed was for Shimmer, followed by Other, HNR, Spectral and Jitter, with MFCCs contributing the least (see Figure 10). At the class level (mean $\pm$ standard deviation), Shriek and Scream showed highly similar profiles, with the highest mean group weights in Shimmer and Jitter, whereas Clear Voice was driven by Spectral and MFCC features. Growl was most strongly associated with Shimmer and showed elevated contributions from Other and HNR, indicating that Growl, unlike the other categories, is mainly characterised by stronger amplitude‑perturbation cues, while Shriek and Scream likely depend on finer‑grained patterns within these perturbation groups.

Figure 10

In line with vocology accounts that extreme vocal techniques arise from non‑traditional vibratory behaviours in the vocal folds and other laryngeal structures (Erbe, 2014; Nieto, 2008; Smialek et al., 2012), the prominence of perturbation‑related groups in our model suggests that irregular vibration and increased noise components provide salient cues for distorted techniques. This is consistent with Kato and Ito (2013), who reported higher Jitter and Shimmer and lower HNR in Growl and screaming voices compared to Clear Voice, highlighting perturbation‑ and harmonicity‑related differences between distorted and non‑distorted phonation.

5.2.3 Human‑prediction task

We trained an additional ComParE SVM to predict human majority decisions. Using the same experimental setup as in the baseline classification task, the model achieved a pooled test performance of 76.1% accuracy and 78.3% UAR (N = 71). Across folds, the highest mean group weight in the human‑prediction model was observed for the MFCC category, followed by HNR, Shimmer, Jitter and Other, with Spectral contributing least (see Figure 11). At the class level (mean ± standard deviation), Shriek was most strongly associated with MFCC and HNR, Clear Voice showed highest contributions from HNR, Growl showed highest contributions from MFCC and Shimmer and Scream relied comparatively more on Jitter and HNR.

Figure 11

5.2.4 Feature group importance comparison

Figure 12 compares the broad feature group profiles between the baseline classification model and the human‑prediction model. While the baseline classification model is dominated by Shimmer (rank 1) and places the MFCC category in rank 6, the human‑prediction model shifts the MFCC category to the top (rank 1) and pushes Spectral to rank 6. This ranking shift in our human‑prediction task suggests that listener majority decisions may place greater weight on MFCC‑based timbral cues, supporting the interpretation that the human prediction model relies more on perceptually salient timbral information and that humans may also place more weight on HNR‑related cues linked to perceived periodicity versus noisiness and voice quality. In contrast, the dataset label classifier is dominated by shimmer‑related cues; relative to the main model, Spectral ranks lowest in the human‑prediction model, and Other features become less prominent in this comparison.

Figure 12

5.2.5 Ablation analysis

We evaluate the hypothesis that the baseline classification model may distinguish Shriek and Scream through finer‑grained combinations within perturbation‑related cue families to a greater extent than through differences between broad feature groups. In our feature group analysis (see Figure 10), Shriek and Scream indeed showed highly similar group level importance profiles, both being dominated by perturbation‑related cues, especially Shimmer and Jitter. By contrast, Growl showed a more distinct pattern, with stronger Shimmer and additional contributions from HNR and Other. Moreover, the feature group analysis in both the human‑prediction task and the baseline classification task showed that HNR ranked second and third, respectively (see Figure 12).

The ablation results further support this interpretation. When only Jitter, Shimmer and HNR features (i.e. the restricted model) were retained, pooled UAR decreased from 90.0% to 76.7% and bidirectional Shriek or Scream confusion increased: Shriek to Scream remained 15.0%, while Scream to Shriek increased from 0.0% to 10.5%; see Figure 13. When Jitter, Shimmer and HNR features were removed (i.e. the complementary model), pooled UAR also decreased to 82.3%, and the opposite confusion direction increased further: Shriek to Scream decreased from 15.0% to 10.0%, while Scream to Shriek increased to 15.8%; see Figure 14. Taken together, these results support the view that the Shriek versus Scream distinction is not primarily based on different broad acoustic cue families but more likely on finer‑grained combinations within and across cue groups, which may also help explain the elevated human confusion between the two categories.

Figure 13

Figure 14