Cross Corpus Speech Emotion Classification - An Effective Transfer Learning Technique

Cross Corpus Speech Emotion Classification - An Effective Transfer Learning Technique Siddique Latif 1,3, Rajib Rana 2, Shahzad Younis 1, Junaid Qadir 3, and Julien Epps 4 1 National University of Sciences and Technology (NUST), Pakistan 2 University of Southern Queensland, Australia 3 Information Technology University (ITU)-Punjab, Pakistan 4 University of New South Wales, Sydney, Australia arxiv:11.6353v2 [cs.cv] 22 Jan 18 Abstract Cross-corpus speech emotion recognition can be a useful transfer learning technique to build a robust speech emotion recognition system by leveraging information from various speech datasets - cross-language and cross-corpus. However, more research needs to be carried out to understand the effective operating scenarios of cross-corpus speech emotion recognition, especially with the utilization of the powerful deep learning techniques. In this paper, we use five different corpora of three different languages to investigate the cross-corpus and cross-language emotion recognition using Deep Belief Networks (s). Experimental results demonstrate that s with generalization power offers better accuracy than a discriminative method based on Sparse Auto Encoder and SVM. Results also suggest that using a large number of languages for training and using a small fraction of target data in training can significantly boost accuracy compared to using the same language for training and testing. I. INTRODUCTION In recent years, speech emotion recognition has received an increasing amount of interest. Speech emotion recognition focus on using linguistic and acoustic attributes as input features and machine learning models as classifiers to predict the emotions of the speaker [1]. These systems achieve promising results when training and testing are performed from the same corpus. However, for real applications, such system does not perform well when speech utterances exist in different languages and collected from different age groups in quite different conditions. At present, various emotional corpora exist, but they are dissimilar based on the spoken language, type of emotion (i.e., naturalistic, elicited, or acted ) and labeling scheme (i.e., dimensional or categorical) [2]. There are more than 5, spoken languages around the world, but only 389 languages account for 94% of the worlds population 1. Even for 389 languages, very few adequate resources (speech corpus) are available for language and speech processing research. It means that the research in language and speech analysis confront with the problem of data scarcity. This imbalance, variation, diversity, and dynamics in speech and language databases make almost impossible to learn a model from a single corpus and put it on the shelf. 1 https://www.ethnologue.com/statistics In automatic speech emotion recognition, most of the studies focus on single corpus without considering the linguistic difference. However, ever since transfer learning has started solving cross-domain classification and pattern recognition problems, interests are growing in cross-corpus emotion recognition. Transfer learning focuses on adapting knowledge from available auxiliary resources to transfer this learning to a target domain, where a very few or even no labeled data is available [3], [4]. Deep neural networks (DNNs) based transfer learning has improved image classification by using a very large dataset as source domain and small data as a target domain [5]. Some studies also exploited DNNs for transfer learning in speech analysis. For example, the performance of sound event classification is improved by training DNNs on very large speech data [6]. In [7], authors used a single DNNs for language and speaker recognition with a large gain on performance by training the model on speech recognition data. Because, most of previous studies do not provide extensive insights on cross-corpus and cross-language speech emotion recognition. We therefore investigate this across multiple languages (English, German, Italian) using different databases. We have used five publicly available databases and performed pairwise experiments using comparatively large dataset as training domain and other small datasets as a target domain to explore the feasibility of transfer learning for various scenarios. We believe that our exploratory research will help researchers gather valuable knowledge about using cross-corpus and crosslanguage speech emotion recognition as a transfer learning technique and apply it in their relevant area of research. II. RELATED WORK Although cross-language and cross-corpus speech emotion recognition is an interesting topic, very few studies have exploited this concept. The existing studies have mostly studied the preliminary feasibility of cross-corpus learning and warranted for further in-depth research. For example, Schuller et al. [2] used six different corpora to analyze cross-corpora emotion recognition using SVM and highlighted the limitation of the current system for cross-corpus emotion recognition. Eyben et al. [8] used four corpora to evaluate some pilot experiments on cross corpus emotion recognition while using

support vector machines (SVM). They used three datasets for training and forth for testing and showed that the cross-corpus emotion recognition is feasible. To explore the universal cue of emotions regardless of language, Xia et al. [9] investigated cross-language emotion recognition for Mandarin vs. Western Languages (i.e., German, and Danish). The authors focused on gender-specific speech emotion recognition and achieved the classification rates higher than the chance level. Albornoz et al. [] developed an ensemble SVM for emotion detection with a focus on emotion recognition in never seen languages. In this paper we use deep leaning within transfer learning for speech emotion recognition. Deep learning based transfer learning has recently been used for speech analysis and recognition. However, the existing research has focused on basic DNNs, but the impact of using complex DNN models suitable for cross-corpus/cross-language setting is yet to be explored. We use s and demonstrate that it performs better than a popular DNN model - autoencoders. The key reason to employ is its power of generalisation, which is not present in most conventional DNN models. Intuitively, for cross-corpus and cross-language emotion recognition, generalisation power of a model is crucial. Example of some existing research are as follows. In [7], authors used a single DNN for speaker and language recognition with a large gain on performance by training the model on speech recognition data. Deng et al. [11] used sparse autoencoders for feature transfer learning in speech emotion recognition. They used six standard databases and use single-layer sparse autoencoder and train this model on classspecific instances from the target domain and then apply this representation to source domain for reconstruction of those data. This experimental approach improves the performance of the model as compared to the independent learning from every source domain. Lim et al. [6] proposed cross-acoustic transfer learning framework by using DNNs. After a series of experiments, the obtained results showed that the crossacoustic transfer learning can significantly enhance the classification rate. Apart from DNNs, researchers have also used interesting deep architectures for transfer learning. In [12], the authors focused on using Progressive Neural Networks to transfer knowledge for three paralinguistic tasks, i.e., emotion, speaker, and gender detection. Progressive Networks are useful for conducting multitasking in a network, however, we focus on a single task of emotion recognition as speaker and gender recognition are not the focus of this paper. A. Speech Databases III. EXPERIMENTAL SETUP Five corpora in three languages are covered in this paper for computational experiments on cross-corpora and crosslanguage emotion detection. These databases are annotated differently, therefore, one of the only consistent ways to investigate transfer learning is by looking at positive/negative valence. We adopt the binary valence mapping per emotion category from [2], [11], [18], [19]. The name of the datasets used in our experiment and the categorical mappings to binary valence are provided in Table I. These databases are chosen to span a wide variety of languages and they are widely used by the researchers. B. Speech Features In this study, we use egemaps feature set that is a widely used benchmark features set for speech emotion recognition studies [12]. The feature set includes Low-Level Descriptor (LLD) features of the speech signal that have been proposed as the most relevant to emotions by Paralinguistic studies [19]. The egemaps feature set contains 88 features including frequency, energy, spectral, cepstral, and dynamic information. The overall components are the arithmetic mean and coefficient of variation of 18 LLDs, 6 temporal features, 4 statistics over the unvoiced segments, 8 functionals applied to loudness and pitch, and 26 additional dynamic and cepstral components. C. Deep belief Networks We have used Deep Belief Networks (s) for emotion recognition from speech. s are very popular deep architectures that consist of the stack of Restricted Boltzmann Machines (RBMs) to make a powerful probabilistic generative model by using layer-wise training in a greedy manner. RBM is an undirected stochastic neural network consisting of a visible layer, a hidden layer, and a bias unit. Each visible unit of the visible layer is fully connected to hidden units in the hidden layer, and the bias is connected to all the visible units and the hidden units. There is no connection between a visible to visible and hidden to hidden units. RBM can also be used as classifiers. It is trained on the joint distribution of input data and corresponding labels, then it assigns the label to new input which has the highest probability under the model. Deep belief networks (s) widely used for speech analysis and they are proved to be very successful for highlevel feature learning and emotion classification [], [21]. In particular, can learn more powerful and effective discriminative long-range of features [22] that are helpful in the speech-related problem like emotion recognition [23]. These abilities of have not been exploited for crosscorpus emotion recognition, therefore, we choose for this task. In this experiment, we use having three RBM layers, where the first two RBMs have hidden unit each, and the third RBM have hidden units with learning rate of 3 and epochs. This implementation is adopted from [24], [25]. We used layer-by-layer pre-training in, and backpropagation technique through the whole network to fine-tune the weights to maximise classification accuracy. IV. EXPERIMENTATION In this section, we explore various scenarios for crosscorpus and cross-language speech emotion recognition and conduct experiments to test the scenarios.

TABLE I: Corpora information and the mapping of class labels onto Negative/Positive valence. Corpus Language Age Utterances Negative Valance Positive Valance References FAU-AIBO German Children 18216 Angry, Touchy, Emphatic, Reprimanding Motherese, Joyful, Neutral, Rest [13] IEMOCAP English Adults 5531 Angry, Sadness Neutral, Happy, Excited [14] EMO-DB German Adults 494 Anger, Sadness, Fear, Disgust, Boredom Neutral, Happiness [15] English Adults 4 Anger, Sadness, Fear, Disgust Neutral, Happiness, Surprise [16] Italian Adults 588 Anger, Sadness, Fear, Disgust Neutral, Joy, Surprise [17] A. Baseline Comparison In order to obtain the baseline comparison results, we compare the performance of with a popular approach of using sparse autoencoder (AE) with SVM [11]. For baseline experiments 75% of randomly selected data is used for training and remaining 25% is used for testing. Figure 1 shows the comparison results, where outperforms sparse AE for all databases. 9 74.11 71.73 54.77.29 72.38.24 56.76.95 76.22 64.76 FAU-AIBO IEMOCAP EMO-DB Fig. 1: Comparison of baseline accuracy using and sparse AE on different databases. B. Language Tests In this experiment, we use one language dataset for training and use the rest of the language datasets for testing. For brevity, we just use FAU-AIBO and IEMOCAP datasets for training. In order to evaluate the model on IEMOCAP, we used two sessions out of five with two fold cross validation because overall data is large. The other databases are small comparative to IEMOCAP, therefore, we used them completely. Figure 2 shows the achieved recognition rate in these experiments and its comparison with previous technique using sparse autoencoder and SVM for cross-corpus transfer learning. When IEMOCAP database is used for training the, we performed pair wise testing by using OHM and MONT separately for FAU-AIBO. It can be noted from Figure 2 that outperforms sparse AE for all scenarios. Beyond this we do not present the accuracy of sparse AE anymore, as we observe that s consistently outperform sparse AE. C. Multi-language Training In this experiment, we use multiple languages jointly for training to observe if that improves the performance of using languages individually for training. We use both FAU-AIBO and IEMOCAP for training and remaining for testing. We also evaluate the model within the corpora. For IEMOCAP, we used three sessions (plus FAU-AIBO) for training and testing is performed using remaining two sessions with two-fold cross validation. Similarly, for FAU-AIBO, a similar two fold crossvalidation is used, i.e., training on OHM (plus IEMOCAP) and evaluating on MONT and the inverse. Further, we also performed training in the leave-one-dataout scheme. For FAU-AIBO, we have performed evaluation by using OHM and MONT independently and take the average results. In case of IEMOCAP, we used two sessions (with two fold cross validation) to evaluate the model. It performs better than baseline and two-language training as shown in Figure 3. D. Percentage of Target Data In this experiment, we use the percentage (% to %) of target dataset for the training of the model. The training was performed using IEMOCAP and FAU-AIBO separately and, EMO-DB and were used for testing. The results have been shown in Figure 4. The straight lines in the figure show the basline recognition rate of respective corpus. These results show that the recognition rate significantly improves (than baseline) by including target domain data with the training data. V. DISCUSSIONS From the experiments, Leave-One-Out seems to be standing out in-terms of obtaining the highest accuracy. This essentially means that training the model using a large range of languages would help learn many intrinsic features from each languages, which can essentially help to achieve high accuracy in an unknown language - even higher than when the same language is used for training and testing (baseline). The performance of the Leave-one-out (see Figure 3) on database is a prime example of this. Both German and English language have two datasets each, i.e., in leave one out scheme there will at least one of these language in the training set. But for there will be a situation that emotions in the Italian language are predicted simply based on emotions in German and English language. Although this is an useful finding, it warrants further investigation to pinpoint the underlying reasons for this improved performance. Another interesting aspect we learned from the experiments that using fraction of the target data into training can help improve the performance and help achieve better results than baseline. Based on our experiments around % of data from the target database can help achieve better than the baseline accuracy. However, this is worse while using FAU-AIBO for training. Interestingly, IEMOCAP performs well on EMO-DB that is in the German language as compared to FAU-AIBO that is also in German. This warrants further investigation if

Accuracy (%) Accuracy (%) EMO-DB (Italian) FAU-AIBO EMO-DB (Itaian) IEMOCAP (a) Recognition rate using IEMOCAP for training and other databases for testing. Fig. 2: Comparison of language tests using and sparse AE. (b) Recognition rate using FAU-AIBO for training and other databases for testing. 9 FAU-AIBO IEMOCAP EMODB Baseline FAU+IEMOCAP Leave One Out Fig. 3: Comparison of baseline results and transfer learning using FAU-AIBO+IEMOCAP and leave-one-language-out scheme. (Italian) this is due to age difference as FAU-AIBO consists of children speech whereas EMO-DB database contains adult speech. Even more interestingly, the language test results in Figure 2 show that performance of IEMOCAP and FAU- AIBO on both and EMO-DB are almost the same. This raise the question if the age factor as discussed before be only applicable while using a fraction of data from the target domain. Finally, based on our experimental results, deep Learning model selection is also important to achieve high accuracy. We observe from Figure 1 and Figure 3 that networks with generalisation power like Deep Belief Networks are more preferable than conventional discriminitive networks like sparse auto encoders. More experiments need to be conducted to genaralise this findings. VI. CONCLUSION We conclude that cross-corpus and cross-language speech emotion recognition is definitely a very useful transfer learning technique. Information from multiple languages or even different datasets in the same language can be leveraged to achieve high accuracy in even an unknown language. For practical applications, this would be very helpful to build a robust speech emotion recognition system using data from multiple languages. Also, this would be equally useful for emotion recognition in languages with very limited or no datasets. REFERENCES [1] A. Batliner, B. Schuller, D. Seppi, S. Steidl, L. Devillers, L. Vidrascu, T. Vogt, V. Aharonson, and N. Amir, The automatic recognition of emotions in speech, in Emotion-Oriented Systems. Springer, 11, pp. 71 99. [2] B. Schuller, B. Vlasenko, F. Eyben, M. Wollmer, A. Stuhlsatz, A. Wendemuth, and G. Rigoll, Cross-corpus acoustic emotion recognition: Variances and strategies, IEEE Transactions on Affective Computing, vol. 1, no. 2, pp. 119 131,. [3] S. J. Pan and Q. Yang, A survey on transfer learning, IEEE Transactions on knowledge and data engineering, vol. 22, no., pp. 1345 1359,. [4] J. Lu, V. Behbood, P. Hao, H. Zuo, S. Xue, and G. Zhang, Transfer learning using computational intelligence: a survey, Knowledge-Based Systems, vol., pp. 14 23, 15. [5] Y. Sawada and K. Kozuka, Transfer learning method using multiprediction deep boltzmann machines for a small scale dataset, in Machine Vision Applications (MVA), 15 14th IAPR International Conference on. IEEE, 15, pp. 1 113. [6] H. Lim, M. J. Kim, and H. Kim, Cross-acoustic transfer learning for sound event classification, in Acoustics, Speech and Signal Processing (ICASSP), 16 IEEE International Conference on. IEEE, 16, pp. 24 28. [7] F. Richardson, D. Reynolds, and N. Dehak, Deep neural network approaches to speaker and language recognition, IEEE Signal Processing Letters, vol. 22, no., pp. 1671 1675, 15. [8] F. Eyben, A. Batliner, B. Schuller, D. Seppi, and S. Steidl, Crosscorpus classification of realistic emotions-some pilot experiments, in Proc. LREC workshop on Emotion Corpora, Valettea, Malta,, pp. 77 82. [9] Z. Xiao, D. Wu, X. Zhang, and Z. Tao, Speech emotion recognition cross language families: Mandarin vs. western languages, in Progress in Informatics and Computing (PIC), 16 International Conference on. IEEE, 16, pp. 253 257. [] E. M. Albornoz and D. H. Milone, Emotion recognition in never-seen languages using a novel ensemble method with emotion profiles, IEEE Transactions on Affective Computing, vol. 8, no. 1, pp. 43 53, 17. [11] J. Deng, Z. Zhang, E. Marchi, and B. Schuller, Sparse autoencoderbased feature transfer learning for speech emotion recognition, in Affective Computing and Intelligent Interaction (ACII), 13 Humaine Association Conference on. IEEE, 13, pp. 511 516. [12] J. Gideon, S. Khorram, Z. Aldeneh, D. Dimitriadis, and E. M. Provost, Progressive neural networks for transfer learning in emotion recognition, arxiv preprint arxiv:16.3256, 17. [13] B. Schuller, S. Steidl, and A. Batliner, The interspeech 9 emotion challenge, in Tenth Annual Conference of the International Speech Communication Association, 9. [14] C. Busso, M. Bulut, C.-C. Lee, A. Kazemzadeh, E. Mower, S. Kim, J. N. Chang, S. Lee, and S. S. Narayanan, Iemocap: Interactive emotional dyadic motion capture database, Language resources and evaluation, vol. 42, no. 4, p. 335, 8.

% Accuracy 85 75 65 55 55 % % % % % % % % % of target date with training data EMODB (a) Training using IEMOCAP. Fig. 4: Impact of using a percentage of target date with training data. % Accuracy 85 75 65 % % % % % % % % EMODB (b) Training using FAU-AIBO. [15] F. Burkhardt, A. Paeschke, M. Rolfes, W. F. Sendlmeier, and B. Weiss, A database of german emotional speech. in Interspeech, vol. 5, 5, pp. 1517 15. [16] P. Jackson and S. Haq, Surrey audio-visual expressed emotion(savee) database, University of Surrey: Guildford, UK, 14. [17] G. Costantini, I. Iaderola, A. Paoloni, and M. Todisco, Emovo corpus: an italian emotional speech database. in LREC, 14, pp. 31 34. [18] H. Sagha, J. Deng, M. Gavryukova, J. Han, and B. Schuller, Cross lingual speech emotion recognition using canonical correlation analysis on principal component subspace, in Acoustics, Speech and Signal Processing (ICASSP), 16 IEEE International Conference on. IEEE, 16, pp. 5 54. [19] F. Eyben, K. R. Scherer, B. W. Schuller, J. Sundberg, E. André, C. Busso, L. Y. Devillers, J. Epps, P. Laukka, S. S. Narayanan et al., The geneva minimalistic acoustic parameter set (gemaps) for voice research and affective computing, IEEE Transactions on Affective Computing, vol. 7, no. 2, pp. 19 2, 16. [] C. Huang, W. Gong, W. Fu, and D. Feng, A research of speech emotion recognition based on deep belief network and svm, Mathematical Problems in Engineering, vol. 14, 14. [21] R. Xia and Y. Liu, A multi-task learning framework for emotion recognition using 2d continuous space, IEEE Transactions on Affective Computing, vol. 8, no. 1, pp. 3 14, 17. [22] G. E. Hinton and R. R. Salakhutdinov, Reducing the dimensionality of data with neural networks, science, vol. 313, no. 5786, pp. 4 7, 6. [23] L. Deng, M. L. Seltzer, D. Yu, A. Acero, A.-r. Mohamed, and G. Hinton, Binary coding of speech spectrograms using a deep auto-encoder, in Eleventh Annual Conference of the International Speech Communication Association,. [24] R. Rana, Emotion classification from noisy speech-a deep learning approach, arxiv preprint arxiv:13.591, 16. [25] M. A. Keyvanrad and M. M. Homayounpour, A brief survey on deep belief networks and introducing a new object oriented toolbox (deebnet), arxiv preprint arxiv:18.3264, 14.