‘Should we laugh?’ Acoustic features of (in)voluntary laughters in spontaneous conversations

However, the previous research managed laughter as one group; Scott et al. (2014) distinguished between two types of laughter. Different research labelled them in different ways: voluntary/involuntary (Scott et al. 2014; Chen 2018) or voluntary/evoked (Scott et al. 2014), spontaneous, authentic/fake (Lavan et al. 2016), social/spontaneous (Shochi et al. 2017), spontaneous/volitional (Bryant et al. 2018; Kamiloğlu et al. 2021), mirthful/polite (Tanaka and Campbell 2014; Sabonyté 2018), authentic/acted (Anikin and Lima 2018)—depending on the framework of the given analyses, in some research used as synonyms. However, the two concepts differed not only in their names, but often in their definitions as well.

Some studies have selected the sound samples for a perception test on the basis of a preliminary grouping:

Similarly, Shochi et al. (2017) investigated the types of laughter in regard to their voluntariness. Firstly, 3 Japanese males and 4 French subjects were asked to listen to 254 laughter that were collected from 12 spontaneous conversations during online video games and decide whether it was social (‘the person is laughing to maintain the communication with the other (e.g. embarrassed laughter, polite laughter, cynical laughter…’) or spontaneous (‘the person is laughing in a spontaneous manner to an external event (e.g. a funny clip)’)). Additionally, ‘I don’t know’ answer was also possible. Then, from altogether, 27 spontaneous and 27 volitional social laughs built up the dataset. In the perception test, 20 Japanese and 82 French native listeners listened to the stimuli and decided what kind of laughter they heard. Results showed that subjects were able to differentiate these types of laughter with about 70% accuracy based only on audio information without their context. Furthermore, acoustic analysis was also conducted regarding these two types of laughs. According to the multiple factor analysis, judgements of both French and Japanese groups were correlated with f0 features (mean and standard deviation), the total duration and the voiced segment duration. Therefore, these acoustic factors were further investigated, and results showed that the total duration of laughs is an important cue for the differentiation regarding their voluntariness. In addition, the voiced duration, the number of voiced segments and the f0 standard deviation also assist in the differentiation between spontaneous and social laughs. The variations of f0 values were higher; the total duration and the voiced segment duration were longer in the case of spontaneous laughs than social ones.

Other examinations have considered the laughs in different (genre) recordings as belonging to one group or the other:

Additionally, Bryant et al. (2018) conducted a perception test on two types of laughter as well. Participants had to decide whether the laugh was real or fake. The test contained 18 spontaneous laughs from natural conversations (real laughter) and 18 ‘fake/volitional’ laughs. The laughter regarding the first category was collected from 13 conversations, involving female friend speakers, while volitional laughter as samples for second category were recorded from women who were instructed to ‘now laugh’ with no other prompting. This task was conducted with 884 participants from six regions of the world. The overall rate of correct judgments was 64%, which was a performance significantly better than the chance level in the differentiation of real and fake laughs. Results showed that people are able to distinguish spontaneous and fake laughs, regardless of their language or culture. Laughs produced with greater intensity variability, higher pitch, and increased noisy features are considered to be spontaneous.

In another study (Kamiloğlu et al. 2021), spontaneous laughs were elicited by funny videos (they laughed in response to self-selected humorous recordings), while participants were instructed to politely laugh at unfunny jokes for collecting volitional laughter from the similar people. In total, almost eight hundred laughter samples were collected from Dutch and Japanese speakers. Then, 20 Dutch and 18 Japanese participants were asked to answer two questions: ‘Do you think this was a genuine or a polite laugh?’ and ‘Did this laugh sound authentic or not?’—they were asked to choose the yes or no response. Sixteen clips (eight Dutch, eight Japanese; laughter type and gender balanced for each group) that were most accurately discriminated as spontaneous versus volitional and that were judged as most authentic were selected as stimuli for the main experiment. Statistical analysis corroborated previous results: spontaneous laughs had higher rates of intervoicing interval, longer duration, increased f0, F1 and F2 means, lower amplitude variability, higher values of spectral centre of gravity and reduced harmonics-to-noise ratios. These 16 laughs were the stimuli of the main experiment. Participants were asked to hear decontextualised laughs and then decide (i) whether it was spontaneous or volitional; (ii) whether the laughing person was from their own or foreign cultural group. Participants also rated the positivity of each stimulus on a 7-point Likert scale. Results showed that both Dutch and Japanese participants rated spontaneous laughter as more positive than volitional ones. However, no difference was found in the accuracy of group membership identification from spontaneous versus volitional laughter.

Other analyses have selected the two different types of laughter from the same social context:

Lavan et al. (2016) conducted an experimental study on the acoustic features and perceptual judgement of volitional and spontaneous laughter. Female speakers were asked to produce both spontaneous (‘genuine amusement laughter’) and volitional (‘voluntary, controlled’) laughter. Then, 72 stimuli were selected for a perception test. Nineteen participants were asked to rate the valence and arousal of the stimuli on a 1–7 Likert scale. According to the acoustic analysis, significant differences were found between volitional and spontaneous laughter for most of the measured acoustic parameters: longer total duration, shorter burst duration, higher f0 mean, higher f0 minimum and maximum, a larger f0 variability, a higher percentage of unvoiced segments and lower mean intensity were measured in the case of spontaneous laughter compared to volitional ones. However, they did not find differences between the two laughter types in f0 range, harmonics-to-noise ratio (HNR) and spectral centre of gravity. Furthermore, the results of the perceptual experiment showed that spontaneous and volitional laughter were perceived as being different in arousal, valence, and authenticity; therefore, participants were able to distinguish between these two types of laughter. Combinations of the laughter’s total duration, spectral centre of gravity, and f0 mean were the most prevalent predictors for ratings of spontaneous laughs. In contrast, in the case of volitional laughs, HNR was found to be the most frequent predictor for affective ratings. Results showed that volitional laughs, with a lower HNR, appeared to be more authentic and more positive.

In another study, 100 samples of non-overlapping spontaneous, mirthful and polite laughter were collected from daily conversations and TV talk shows (Sabonytė 2018). Thirty stimuli were used; 30 respondents were asked to label the type of laughter after hearing the recordings with and without the context. Acoustic features of the two types of laughter were analysed. Data showed significant difference in duration between mirthful and polite laughs, but neither in intensity, in f0, F1, F2, in shimmer and jitter values. Bouts of mirthful laughter were longer than bouts of polite laughter of the same number of structures: polite laughter contained one, whereas mirthful laughter consisted of one and more bouts. In the case of polite laughter, the most common form was a two-syllable laughter bout, the mirthful laughter may consisted of more than one bout; the bouts of this type of laughter consisted of one to fifteen syllables (the most common samples are of three or four syllables). In addition, differences were found between the accuracy of cluster analysis and human perception in distinction of these laughter types.

Differentiation between acted and authentic emotional non-verbal vocalisation was analysed (Anikin and Lima 2018). The judgement task was conducted online. The participants were asked to listen to various sounds (from seven different corpora) and decide whether they were real (authentic) or fake (pretending). Results showed that the accuracy of authenticity detection varied regarding the emotional category. Authentic non-verbal vocalisations of fear, anger and pleasure were much more likely to be deemed as authentic as posed vocalisations. The authentic laughs were perceived as authentic in 67% of all cases, and they occurred with higher pitch, larger pitch variability and lower harmonicity than fake ones.

Other studies also investigated the influence of the relation of the speakers on laughter perception (Farley et al. 2022). Laughter stimuli were obtained from telephone calls of 27 callers talking to their romantic partner and a close same-sex friend. Fifty-two samples were selected for the study: in the first task, these laughter had to be judged by the 50 raters regarding pleasantness. In the second task, listeners had to decide whether the laughter was directed towards a friend or a romantic partner. Results showed that listeners were able to identify them in a higher proportion than the chance level (57%). Furthermore, laughter directed at romantic partners were judged to be less pleasant-sounding than those directed at friends. Additionally, in the second part of the study, the eight, most prototypical laughs were selected. Participants had to judge laughter samples regarding spontaneity using bipolar scales (e.g. ‘loud/soft’, ‘natural/forced’, ‘breathy/not breathy’) and regarding vulnerability. Laughter directed at friends were judged to be louder, more masculine, natural-sounding, ‘changing’, mature-sounding, more dominant, and less breathy than those directed at romantic partners. The gender of the speakers also affected judgements significantly: laughter samples from male speakers received higher ratings for masculinity and coldness, while female laughter samples were higher rated for loudness, naturalness, changing, maturity, relaxed, and dominance.

Beside the research focusing on the judgements of laughter types, another group of research aimed at the automatic classification of different kinds of laughs. The laughter detector developed by Campbell et al. (2005) can automatically recognise four laughter types based on the speaker’s affective state and their segmental composition (voiced laugh, chuckle, ingressive breathy laugh, nasal grunt). The identification rate was greater than 75%. In Galvan et al. (2011), automatic recognition based on vocal features also achieved high accuracy scores (70% correct recognition) when discriminating five types of acted laughter: happiness, giddiness, excitement, embarrassment and hurtful.

By social function, samples of laughter were divided into five groups: mirthful, politeness, embarrassment, derision and others (Tanaka and Campbell 2011). In natural communication, the most frequent types of social laughter seemed to be polite and mirthful (Tanaka and Campbell 2011, 2014; Sabonytė 2018). In order to distinguish between different types of laughter, and based on phonetic characteristics (voiced, ingressive, chuckle, nasal), Tanaka and Campbell (2011) used HMM with the following spectral features: MFCC, RMS power, and delta, power, and achieved a prediction accuracy of 86.79%.

In another study, Tanaka and Campbell (2014) categorised laughs into either polite or genuinely mirthful categories (based on the majority vote of 20 observers). They determine the main contributing factors in each case by statistical analysis of the acoustic features, principal component analysis and classification tree analysis. SVM was used to predict the most likely category for each laugh in both speaker-specific and speaker-independent manner. Better than 70% accuracy was obtained in automatic classification tests.

Through the investigation of laughter-related body movements, five laughter states (hilarious, social, awkward, fake, and non-laughter) were distinguished automatically by Griffin et al. (2015).

The automatic detection and classification of laughter occurrences can be beneficial in a number of ways. It could be used in automatic speech recognition (ASR) systems, reducing the word error rate by identifying non-speech sounds. It can be helpful in searching for videos with humorous content. Detection of users’ emotional state from various modalities (body movements, facial expressions, speech) and production of emotional displays can be used in design of human–computer interaction (HCI). Automatic detection of laughter can be useful for detecting the user's affective state and conversational signals such as agreement. Thus, it may facilitate affect-sensitive multimodal human–computer interfaces. Virtual/embodied agents could be made more natural (human-like) using natural-sounding synthesised laughter.

Previous research has therefore made several findings regarding the categorisation of laughter. The previous research examined the different realisations of laughter from many aspects: in the introduction part, we focused on the results obtained based on three main aspects: the production, the perception and automatic categorisation and classification. However, drawing general conclusions is made more difficult by the fact that the individual studies differed in many respects: Some research contrasted spontaneous laughter with fake (Lavan et al. 2016), social (Shochi et al. 2017), volitional (Kamiloğlu et al. 2021), while other examinations distinguished between voluntary/involuntary (Scott et al. 2014; Chen 2018), voluntary/evoked (Scott et al. 2014), or mirthful/polite (Tanaka and Campbell 2014; Sabonyté 2018). On the one hand, these researches defined the groups of laughter differently. On the other hand—as well in close connection with this fact—they also use different methodologies with regard to the data collection: real laughs were collected from ‘natural, spontaneous’ recordings (see Shochi et al. 2017; Bryant et al. 2018; Sabonytė 2018) or during funny videos (Kamiloğlu et al. 2021), ‘fake’ laughter were forced in an artificial way, for example, they asked the speakers to show how would they laugh at an unfunny joke in a polite way (Kamiloğlu et al. 2021), professional actors were requested to produce different types of laughter (Szameitat et al. 2009), or participants were called just to laugh (Bryant et al. 2018), while in other studies both of the laughter were recorded during video games. In addition, differences have been found in respect with, for example, the analysed acoustic features, measurement methods, as well. The results and the conclusions that can be drawn from them are therefore difficult to generalise and are highly limited: The results of the production and perception examination listed above—independently from the categories used to concepts, methods and/or definitions of the different kinds of laughter—found differences between their categories’ laughs acoustically and perceptually. Most of the research maintained the effect of the timing features regarding that the spontaneous laughter realised with longer duration than the samples in the other category. Although, in the case of f0 characteristics, not all research has corroborated a difference between the two groups, if they do, the mean f0 was higher in the case of the spontaneous or real laughter than as for the other category. Regarding the other parameters, such as the intensity, CoG, F1, F2, pitch, jitter, shimmer, HNR, the results of the research were often contradictory.

The perception test was conducted online. Although the issue often arises that the conditions of the perceptual test are uncontrollable in digital form, Horton et al. (2011) and Germine et al. (2012) results showed that there is no significant difference when the perception test is performed online or in a laboratory. The perception test was developed in the GMS framework system especially for this study. GMS was originally a gamification content management system with numerous game-based learning projects. Now it is also used for personal tests, scientific research and assessment as well. It is completely web-based, responsive and mobile-friendly, using only HTML, CSS and JavaScript code.

Forty-nine laughter stimuli were presented in a random order using an experiment software developed for this study. Participants were asked to use headphones during the test. After listening to two trial stimuli for checking the appropriate volume, they were asked to listen to each laughter and decide whether it was a voluntary or an involuntary one, clicking on the appropriate answer. In each case, they had to choose between the two options, they did not have the option to give an ‘I don’t know’ answer. The voluntary and involuntary options were labelled with text and with signs as well (with a heart and a brain motive, Fig. 1). The procedure used is similar to the judgement task of Bryant et al. (2018); however, the types of the analysed laughs were different. Participants in our task had to decide whether the particular sample seemed to be rather involuntary or voluntary, therefore, no ‘good’ or ‘bad’ answer existed, the acoustic analysis of laughs regarding the two types was based on the judgement of the participants’ majority. The reaction time of the answers was also measured by the given software to see how certain or uncertain the categorisation of the laughter was (in the case of a fast response, the confidence of the response; in the case of a slow response, the uncertainty of the response).

One hundred and seventy people participated in the online perception test, with a mean age of 35.1 years (median 36 years, SD: 11.8 years, min.: 18 years, max.: 67 years): 136 females and 34 males. The degree of the participants varied from elementary school to master degree; however, 50% of the participants had university degrees, 29% had a high school diploma, 16% had completed post-graduate education, 2% had completed technical or vocational education and 3% had only completed primary school. The test participants were all native Hungarian speakers with no known language or hearing problems. Participation in the test was volunteer, no fee was paid. No special features were required for the participation.

For the acoustic analysis, sound samples of laughter were chosen which were considered to be voluntary and involuntary by at least 66.6% of listeners (cf. Shochi et al. 2017). These laughs were analysed and compared with regard to their voluntariness according to the following acoustical features:

Nonparametric Mann–Whitney tests were used for comparing these acoustic parameters of laughter and reaction time of the answering regarding voluntariness. Pearson correlation was adopted for the correlation analysis (1) between the total duration of the given laughter and the ratio of the voiced parts in the laughs and (2) between the acoustic parameters and the proportion of perceptions. Data were analysed by the R program (R Core Team 2018). We examined whether changes in each acoustic parameter were associated with a higher proportion of ratings of voluntary or involuntary for a given laughter sound pattern.

In this study, logistic regression was used to classify laughter types (voluntary or involuntary). The latter was chosen because it is a simple algorithm for binary classification.

We applied this technique on the dataset that was extracted from the perception test. These laughter samples were categorised consistently by human listeners either as voluntary or involuntary laughter. Furthermore, we intended to improve our data source by complementing it with more samples. It is known that human-like abilities can be imparted to machines for specific tasks by means of machine learning algorithms. A classifier and semi-supervised learning algorithm were proposed that make effective use of unlabelled data to improve classification performance (Liang et al. 2007). Given the existence of labels for the laughs from the perception experiment, we applied the label propagation method to a randomly selected larger set of unlabelled laugh samples in order to label them as either voluntary or involuntary. This technique allows us to incorporate not only the knowledge of the labelled data, but also the features of the unlabelled data.

The unlabelled dataset contained 159 manually segmented laughter instances from the Hungarian Spontaneous Speech Database (Neuberger et al. 2014). The same acoustic parameters were evaluated as in the acoustic study, and these features were used for the label propagation procedure: duration, f0 mean, f0 std, HNR, local jitter, local shimmer.

Label propagation and logistic regression were carried out in Python 3.8.10 using Scikit-learn 1.0.2 (Pedregosa et al. 2011).

To measure the advantage of semi-supervised learning with logistic regression, a baseline with only logistic regression without additional methods was run. The labelled data was divided (labelled as either voluntary or involuntary by human listeners in the perception test) into training and test sets (Fig. 2) using stratified split method (i.e. the folds were made by preserving the percentage of samples for each class). The train–test ratio was 70–30% of the entire dataset. Features of the train set were normalised using MinMaxScaler, which scales each feature to a given range, in this case between zero and one.

Baseline logistic regression model: First, we used a logistic regression algorithm fit only on the labelled portion of the normalised training dataset and applied the trained logistic regression model to the normalised test set to establish a baseline in performance on the semi-supervised learning dataset.

Logistic regression model with label propagation: Next, the label propagation method was employed on a dataset consisting of the normalised train set and the normalised unlabelled dataset. Labels were propagated with a combination of random walk and clamping in this algorithm (Zhu and Ghahramani 2002). The goal of the label propagation technique was to predict the labels of the unlabelled samples. We selected samples in which the prediction related to the cases of confidence was higher than 0.8 and added them to a mixed training set together with the normalised train set to train logistic regression classification models on it.

Participants considered almost half of the laughter (45%) uniformly as voluntary/involuntary; however, the other half (55%) were deemed as a compound (the proportion of uniform decisions did not reach 66.6%). From those laughter that were determined as voluntary or involuntary, 46.5% were considered as to be voluntary, while 53.5% of them as to be involuntary (Fig. 3). However, the ratio of the judgements showed great idiosyncrasy across sound samples. Fifty-two percentage of all laughter samples were deemed unanimously by 50–66% of all participants, while 38% of the samples by 67–80% of the participants. Only 10% of the laughs were considered uniformly regarding voluntariness by at least 80% of the listeners.

Reaction time was analysed regarding the results of the judgmental task. Normalised reaction time was analysed because of the participants' different tempo. (Z-score normalisation was applied for each participant.) Considerable difference across the three categories was not found (Fig. 4), although the mean value of the normalised reaction time in the compound category was a bit higher (+ 0.019 ± 1.049) than in the other two categories (involuntary: − 0.014 ± 0.880, voluntary:–0.049 ± 0.984). The boxplot diagram and the standard deviation values also support that there is no considerable difference in the normalised reaction time across the three categories. Normalised reaction time showed great variability across laughter samples as well: both the shortest ( − 1.730) and longest (5.680) normalised reaction times occurred in the compound category. The difference between the groups was not significant.

At first, the acoustic analysis was carried out to evaluate laughter, which were categorised to be voluntary/involuntary by at least 66.6% of all participants (c.f. Shochi et al. 2017).

The mean duration of laughs was 784 ms (SD: 434 ms). The two types of laughter can be well separated based on their duration (Fig. 5). Laughter, which was judged to be involuntary by the majority of the listeners, was longer (mean: 974 ms, SD: 429 ms) than laughter categorised to be voluntary (520 ms, SD: 407 ms). The difference was significant in the duration values between the voluntary and involuntary laughs (Z = − 2.117; p = 0.034).

The mean number of syllables (bursts) in the laughter was the same in the case of involuntary and voluntary laughs (3.4 items); however, involuntary laughter occurred with greater standard deviation (1.6 items) than voluntary ones (1.2 items).

The ratio of the voiced parts of the total duration was also analysed in the case of voluntary and involuntary laughs, categorised uniformly by the majority of the listeners. The voluntary laughter was characterised by a larger standard deviation of this value (mean: 63%, SD: 42%); however, no significant difference was found between laughter types according to the ratio of voiced parts (mean: 49%, SD: 9%). (Fig. 6).

There was also a clear correlation (r = 0.62, p = 0.04) between the duration of the voiced parts and the duration of the laughter: the longer the duration of the laughter was, the greater the proportion of voicedness was (Fig. 7).

Fundamental frequency was higher (mean: 352 Hz) in the case of laughter that were determined to be involuntary by the majority of the participants, than the voluntary ones (mean: 288 Hz). The difference between the two laughter categories was significant: Z = – 2.331; p = 0.020 (Fig. 8).

This difference was found not only in the mean values of f0, but also in the f0 range (Fig. 6). The mean of the f0 range was 4.1 in the case of voluntary laughter (SD: 4.9), while 12.6 of involuntary ones (SD: 6.0). Involuntary laughter was characterised by significantly greater f0 variability (Z = − 3.087; p = 0.002) than voluntary laughter. There was a positive correlation between the timing characteristics and the f0 values in both groups: the listeners considered the laughter with longer duration (r = 0.52, p = 0.019) and higher f0 to be more involuntary (r = 0.47, p = 0.031).

The HNR values were analysed and compared to the case of voluntary/involuntary laughs (categorised uniformly by at least 66.6% of all participants). The HNR values were somewhat higher in laughs that were considered to be voluntary than in the case of involuntary ones; the difference was not significant (Fig. 9). The mean of HNR was 11.03 dB (SD: 4.74 dB) at voluntary laughs, while 6.24 dB (SD: 4.22 dB) at involuntary ones.

Lower jitter values were found; more regular vocal vibrations were considered in laughs that were categorised as voluntary (mean 1.18% SD: 0.07%) than in the case of involuntary ones (mean 1.22% SD: 0.04%). However, statistical analysis did not show any significant difference between the two groups Moreover, in this respect, shimmer showed no difference between laughter deemed to be involuntary (mean 2.37 SD: 0.2%) and voluntary (mean 2.43% SD: 0.2%, see Fig. 10).

The mean centre of gravity (CoG) was 1048 (SD: 540) Hz in involuntary laughter and 517 (SD: 167) in voluntary laughter. The difference between the two laughter types was significant (Z = − 2.385; p = 0.017 (Fig. 11).

Finally, we measured the intensity of voluntary and involuntary laughter (Fig. 12). Involuntary laughter showed lower-intensity values, on average (mean: 64.8, SD: 5.9 dB), than voluntary laughter (mean: 70.4, SD: 5.6 dB). The two laughter types differed significantly in this feature as well (Z = − 2.206; p = 0.027).

Summarising the acoustic features of laughs considered to be voluntary/involuntary by the majority of the listeners; involuntary ones occurred with significantly longer duration, higher F0 mean, higher F0 variability than voluntary ones (Table 1).

Acoustic analysis was carried out on laughter which were not categorised uniformly by the listeners with regard to voluntariness (compound category) as well. The mean duration was 795 ms (SD: 423 ms) of laughter that was longer than the voluntary, and it was shorter than the involuntary ones. However, the difference was not significant.

The f0 parameters of laughter belonging to this compound category were also analysed. Data showed that the f0 mean of these laughter was very similar to laughter considered to be voluntary (289 Hz, SD: 76 Hz) by the majority of the participants. However, it was significantly lower than in the case of involuntary ones (Z = − 2.342; p = 0.019).

Similar to the duration values, the f0 range of laughter of the compound category occurred between the values of voluntary and involuntary categories (mean: 8.7 semitones, SD: 6.1 semitones). The difference was significant between the values of voluntary laughs and the laughs in the compound category (Z = − 2.225; p = 0.024).

Some parameters of voice quality were also analysed in the case of laughter, whose judgement was not so unanimous by the participants. The HNR values in the compound category occurred between the values of voluntary and involuntary laughter (mean: 8.8 SD: 4.8), similar to the duration and f0 range values. However, the jitter values were the highest in the compound category (mean: 1.24% SD: 0.04%) as well as the shimmer values (mean: 2.46% SD: 0. 3%).

Two logistic regression models were built: one of them was trained only on the labelled laughter samples (baseline); the other was trained on labelled and unlabelled data as well (with label propagation). The results demonstrated that both models performed properly. An accuracy of greater than 50% was obtained in 93.5% of cases (187 of 200 runs) for the baseline model and in 98% of cases (196 of 200 runs) for the model with label propagation. In both cases, the performance was significantly greater than 50% (baseline model: t = 18.466; df = 199; p < 0.001; label-prop model: t = 27.257; df = 199; p < 0.001). The models did not show remarkable bias towards classifying laughs as voluntary or involuntary. It is worth noting that both models had a wide range of accuracy. The performance of the logistic regression classification (Fig. 13) showed an improvement over the baseline when label propagation was used. The mean accuracy (Table 2) of the baseline model relying only on human-labelled laughter was 66.14%, whilst the model incorporating a larger set of unlabelled laughter was found to be 73.36% accurate, on average. This indicates that a 10.9% improvement could be achieved in this task using a semi-supervised technique. There was a significant difference between the performance of the two models (t = 6.159; df = 398; p < 0.001). The classification accuracy at the quartiles of the distribution showed a displacement to higher values (the accuracy scores are skewed towards the right side) for the model trained on a larger dataset compared to the baseline. It can also be seen in the distributions that the worst performance yielded the same values of accuracy in both tasks: 42.86%. The best performance of the baseline model yielded an accuracy of 85.71%, while the maximum value of 100% could be achieved in the model with label propagation.

Finally, we checked whether the label propagation works well on a piece of the additional dataset. The number of additional samples was increased along the powers of 2. That means, in the first case, we only added 2 unlabelled laughter to the train set. Then, we continued with the sample addition up to 128. The final train set contained all the 159 unlabelled laughter. The mean accuracy results of each logistic regression model showed a stronger improvement in performance when 16 samples were added and then in the transition from 128 to 159 (Fig. 14).