Humans are emotional creatures. People desire reaction from others according to their emotion . Artificial intelligence (AI) system will be more like human beings when it is able to do such reaction. This capability relies on recognizing emotion. Emotion recognition is therefore very important in AI product, since it will make human-computer interface (HCI) more friendly and improve the user experience.
Emotion recognition system based on audio (which can also be seen as voice-based) has very low requirement for hardware, even though multimodal speech processing can improve speech related system performance [2, 3]. Therefore the audio based emotion recognition system is easier to be employed on AI product [1, 4] than other means. However, current voice-based AI products, such as Siri, Google voice search and Cortona, lack of emotion recognition capability, which make people feel them as “machine”. This shows the importance of exploring on the emotion recognition system.
. So far, most of the work has been done on the data collected in the studio environment. The data collection was well controlled, therefore the data is clean and well segmented. Besides, most of the voice-based emotion recognition research have been done targeting on speech. In real application, there are several problems that may make current developed emotion recognition system failed. First, the non-speech voice fillers, such as laugh, whimper, cry, sigh, sob and etc., has no lexicon information but contains emotion information. Sometimes people only perform non-speech voice filler to express their emotion. Second, the voice segment length may vary in a large range. Conventional feature extraction may fail under this condition. Third, some people can control their intonation and only use the lexical information to express their emotion. Acoustic feature will not work in this case.
To address these problems, we propose an ensemble framework that combines different aspects of features from audio to develop an emotion recognition system applicable in real world. The framework is evaluated on the multimodal emotion challenge (MEC) 2017 corpus. In this study, we focus on categorical emotion recognition, which is the task defined in MEC 2017. The corpus was collected by capturing clips from films and TV programs. These clips may contain background noise and only have non-speech voice filler, which is very close to real world scenarios. The rest parts of the paper are organized as following: Section 2 reviews related work about emotion recognition and previous work on audio-based approach; Section 3 describes the MEC 2017 corpus; Section 4 shows our proposed approach including the sub-system and the ensemble framework; Section 5 shows the experiments results and discuss about results analysis; Section 6 concludes our work and discusses the future work.
2 Related Work
Voice-based emotion recognition has been done for decades. 
extracted prosodic features from speech and applied majority voting of subspace specialists. It was a pilot study exploring static classifier and features for speech-based emotion recognition. built phoneme-based dependent hidden Markov model (HMM) classifier for emotion recognition. This work indicated the speech contents was related to emotion. Both of [6, 4] showed the advantage of HMM over static model.  discussed the feature set for emotion recognition task. The feature sets proposed by these works showed reliable performance.  also used lexical information besides acoustic and showed it was helpful for acoustic event identification. However, most of the work only focused on speech part rather than non-speech voice fillers.
Deep learning techniques were emerging as new classifier in speech related machine learning area. The deep learning techniques, such as
deep neural network(DNN), convolutional neural network (CNN) and recurrent neural network (RNN) is able to model the feature in a high dimensional manifold space. DNN as static classifier and RNN as dynamic classifier showed their advantage in emotion recognition task compared with conventional approaches . Especially,  used attention based weighted pooling to extract acoustic representation. It showed advantage over the conventional hand crafted features. Multi-task learning recently raised as a technique helping train better model for main task [12, 13]. However these work used valence and arousal as auxiliary tasks which may be difficult to get.
In this study, we use deep learning techniques with multi-task learning to build better classifier for categorical emotion recognition. The work includes the following novelties. 1. We combine different features including classical hand-craft feature and high level representation learned from deep learning techniques. The framework considered both speech and non-speech audio. 2. multi-task learning techniques were applied to DNN and weighted pooling RNN with the auxiliary tasks of speaker and gender classification, whose labels can be easily acquired. 3. Lexical information from speech was also incorporated into the system. 4. the framework was targeting at the corpus which is close to real world scenarios.
3 Corpus Description
This study uses the MEC 2017 corpus . The corpus includes the clips from Chinese films, TV plays and talk shows. The clip has both of audio and video. The video is under resolution of 1028 680 with 24 frames per second (FPS). The audio is under 44.1 kHz sample rate with mono channel. In this study, we downsample the audios to 16 kHz. The total duration is 5.6 hours. Each clip has one label among eight classes, angry, worried, anxious, neutral, disgust, surprise, sad and happy. The duration of the clips vary from 0.24 secs to 46.71 secs. The average duration is around 4.1 secs. There are 2105 speakers in this corpus. The gender is almost balanced. The male to female ratio is 0.46 to 0.54. The signal-to-noise ratio (SNR) distribution is shown in Figure 1. The clips is not captured in the controlled studio environment, so there might be background noise in the clips (several clips have lower than 10 db SNR).
Since this is a challenge task, the training and testing set has been determined. The statistics for each category is listed in the Table 1. It can be seen that the data is imbalance but the distribution in training and testing sets are very similar.
4 Proposed Approach
We propose 4 sub-systems in this ensemble framework. The decisions are fused with linear combination. We use the open source toolkit Focal  to fuse the decisions from the 4 sub-systems. It determines the linear combination weight by minimizing cost of the log-likelihood ratio (). The framework diagram is shown in Figure 2
4.1 Multi-task DNN
We built multi-task DNN with the main task of emotion classification and the auxiliary tasks of speaker and gender classification. The system diagram is shown in Figure 3. The assumption that emotion is related to speaker and gender inspires us to incorporate speaker and gender information into the classifier.
The feature set provided in Interspeech 2010 paralinguistic challenge (we name it as “IS10” feature set)  has been used and proved to work well in emotion recognition and speaker ID tasks. Therefore, we select this as one feature set to multi-task DNN. We use openSMILE  to extract the IS10 feature set. iVector  has been proved to work well in speaker ID task. It was also used in emotion classification . Compared with IS10 feature set, it is a high level feature, which contains speaker and channel information. It is also selected as input feature to another multi-task DNN. An iVector extractor is trained based on 2000 hours of cellphone data . For each utterance, a 200-dimension iVector is extracted. iVector has a disadvantage that it may not be reliable with short utterance, while IS10 is not affected by utterance duration. We expect the iVector and IS10 feature set can complement each other.
In this study, we define the architectures of the multi-task DNNs according to the inputs, since they have different dimension number. For the IS10 input, the multi-task DNN has two hidden layers in the trunk part with 4096 neurons in each layer. In the branch part, there are one hidden layer for each task with 2048 neurons. On top of that, there is a softmax layer for task classification. For the iVector input, the architecture is same. But in the trunk part the neuron number is 1024 per layer and in the branch part the neuron number is 1024 per layer. All the neuron type isrectified linear unit
(RELU). The dropout rate is 0.5.
4.2 Lexical Support Vector Machine
We also built a sub-system which is a support vector machine (SVM) based on lexical information. An automatic speech recognition (ASR) system trained with 4000 hours data was used to recognize the speech contents. The ASR is a HMM framework with time-delay DNN (TDNN) acoustic model . After recognizing text, LibShortText toolkit is adopted for text based emotion classification  with the feature of term frequency inverse document frequency
(TF-IDF). The classification is based on support vector classification by Crammer and Singer.
4.3 Attention Based Weighted Pooling RNN
There are utterances in the corpus only containing non-speech voice filler such as laugher, whimper, sob and etc. In additional, the utterance may contain long silence or pause. Using IS10 feature set may not accurately represent acoustic characteristics. In this study, we build a sub-system which is a RNN with attention based weighted pooling to address these issues. RNN was used to model acoustic event [22, 23], which can be voice filler. Attention based weighted pooling has been proved to work better than basic statistics, like averaging, summation and so on, because it is able to capture the informative section rather than silence or pause part . The system diagram is shown in Figure 4.
The input feature is a 36D sequential acoustic feature. The acoustic feature includes 13D MFCCs, zero crossing rate (ZCR), energy, entropy of energy, spectral centroid, spectral spread, spectral entropy, spectral flux, spectral rolloff, 12D chroma vector, chroma deviation, harmonic ratio and pitch. It is extracted from a 25 ms window. The window shifting step size is 10 ms.
The weighted pooling is obtained by the Equation 1, where is the hidden value output from the long short-term memory (LSTM) layer at time , and is the weight. is from time step to (sequence length is n). is a scalar computed by Equation 2, where is the weight to be learned. Equation 2 is a softmax-like equation which is similar to attention model [24, 11]. By learning , it is expected that segments of interest are assigned high weight, while segments of silence or pause are assigned low weight. This model is targeting at not only speech but also other acoustic event, like voice fillers. In this model, we also use multi-task learning, which is expected to learn better representation from weighted pooling compared with single task.
In this study, we used a new type of LSTM, which is named as advanced LSTM (A-LSTM). It is verified that it has better capability of modeling timing dependency. The A-LSTM architecture is proposed in .
This network has two hidden layers in the trunk part. The first one is fully connected layer with 256 RELU neurons. The second one is a bidirectional LSTM (BLSTM) layer with 128 neurons. Weighted pooling is performed on top of the LSTM layer. The representation from weighted pooling is then sent to the branch part. In the branch part, each task has one fully connected layer with 256 RELU neurons. On top of that, there is a softmax layer performing classification. The dropout rate is 0.5.
5 Experiments and Results
The proposed approach is evaluated on MEC 2017 corpus. We build two baseline systems for comparison. One is random forest and the other one is a DNN with single task of emotion classification. Both of the baseline systems take IS10 feature set. The details of experiment settings and results are described in the following sections.
5.1 Experiments Setting
The training data is 4.4 hours and testing data is 0.7 hour. The batch size for all the neural network training in this study is 32. The multi-task DNN was trained with SGD optimizer, while the RNN was trained with Adam optimizer. The task weights for emotion, speaker, gender classification were 1, 0.3, 0.6 respectively. The baseline DNN with single task has three hidden layers. There are 4096 RELU neurons in the first two, and 2048 RELU neurons in the last one. This is the same architecture as multi-task DNN except that it does not have the auxiliary tasks part. The baseline random forest has 100 trees with 10 depth. This is the setting provided by the challenge organizer 
. The IS10 feature was z-normalized with the mean and variance of the training data. The sequential feature was z-normalized within utterance. During training, 10% of training data was set as validation data. The linear combination weight used in fusion was also trained from the validation set.
In the evaluation, we use the metrics of macro average F-score
(MAF) (also called unweighted F-score) and accuracy. The MAF is computed by averaging the F-score for each class detection. The accuracy is computed by dividing correctly detected sample number divided with the total sample number. Since the data is imbalance, accuracy may not be accurate to represent the system performance and MAF will be adopted as the main metric.
5.2 Evaluation Results
The evaluation results are shown in Table 2. The performance of baseline systems, sub-systems and proposed framework is listed.
|System Category||Approach||MAF||Accuracy (%)|
|Sub-system||Multi-task DNN (IS10)||32.4||44.1|
|Multi-task DNN (iVector)||27.4||38.0|
|Weighted Pooling RNN||23.2||39.7|
Comparing the baseline systems, it shows the DNN outperforms the random forest by 4.1% (absolute difference). This shows DNN has better modeling capability even with 4.4 hours training data. The multi-task DNN taking IS10 feature has 6% absolute improvement compared with the baseline DNN. It indicates that the speaker and gender classification tasks are helpful for emotion classification. The multi-task DNN taking iVector feature also outperforms the baseline DNN with 1% absolute gain. This proves iVector can also work well in emotion classification task. The performance of lexical SVM and weighted pooling RNN is lower than the baseline DNN. For the lexical SVM, it relies on the text information from ASR which may not be perfectly reliable. For weighted pooling RNN, the shortage of training data is a key issue. 4.5 hours training to train RNN may not be sufficient. The performance from fusion achieve highest MAF. It shows the sub-systems in the ensemble framework are complementing each other. Compared with the best baseline system, which is the single task DNN, the ensemble framework offers 7.8% absolute improvement (about 29.5% relative improvement).
6 Conclusion and Future Work
In this study, we proposed an ensemble framework for categorical emotion recognition. The proposed framework was evaluated on MEC 2017 corpus, whose data was close to real world scenarios. We found multi-task learning with auxiliary tasks of speaker and gender classification was helpful for emotion classification. Labels for these tasks are normally easily obtained. Fusion of different sub-systems achieved better performance. It indicates capturing different aspects of features from input audio can improve the modeling capability. Since the evaluation was done on acted data in this paper, the proposed framework need be evaluated on the data from real world in the future. Besides, more data is needed for training, which may lead to better performance.
-  C. Busso, Z. Deng, S. Yildirim, M. Bulut, C.M. Lee, A. Kazemzadeh, S. Lee, U. Neumann, and S. Narayanan, “Analysis of emotion recognition using facial expressions, speech and multimodal information,” in Sixth International Conference on Multimodal Interfaces ICMI 2004, State College, PA, October 2004, pp. 205–211, ACM Press.
-  F. Tao and C. Busso, “Lipreading approach for isolated digits recognition under whisper and neutral speech.,” in INTERSPEECH 2014, Singapore, Sep. 2014, pp. 1154–1158.
-  F. Tao and C. Busso, “Bimodal recurrent neural network for audiovisual voice activity detection,” in Interspeech 2017, Stockholm, Sweden, Sep. 2017, pp. 1938–1942.
-  B. Schuller, G. Rigoll, and M. Lang, “Hidden Markov model-based speech emotion recognition,” in ICASSP 2003, Hong Kong, China, April 2003, vol. 2, pp. 1–4.
-  F. Dellaert and T. Polzin A. Waibel, “Recognizing emotion in speech,” in (ICSLP 1996, Philadelphia, PA, USA, October 1996, vol. 3, pp. 1970–1973.
-  C.M. Lee, S. Yildirim, M. Bulut, A. Kazemzadeh, C. Busso, Z. Deng, S. Lee, and S.S. Narayanan, “Emotion recognition based on phoneme classes,” in ICSLP 2004, Jeju Island, Korea, October 2004, pp. 889–892.
-  J. Deng, X. Xu, Z. Zhang, S. Frühholz, D. Grandjean, and B. Schuller, “Fisher kernels on phase-based features for speech emotion recognition,” in Dialogues with Social Robots, pp. 195–203. Springer, 2017.
-  B. Schuller, S. Steidl, A. Batliner, F. Burkhardt, L. Devillers, C. Muller, and S. Narayanan, “The INTERSPEECH 2010 paralinguistic challenge,” in Interspeech 2010, Makuhari, Japan, September 2010, pp. 2794–2797.
-  Q. Jin, C. Li, S. Chen, and H. Wu, “Speech emotion recognition with acoustic and lexical features,” in ICASSP 2015, Queensland, Australia, Apr. 2015, IEEE, pp. 4749–4753.
-  G. Hinton, S. Osindero, and Y. Teh, “A fast learning algorithm for deep belief nets,” Neural computation, vol. 18, no. 7, pp. 1527–1554, 2006.
-  S. Mirsamadi, E. Barsoum, and C. Zhang, “Automatic speech emotion recognition using recurrent neural networks with local attention,” in ICASSP 2017, New Orleans, U.S.A., Mar. 2017, IEEE, pp. 2227–2231.
-  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, 2017.
-  S. Parthasarathy and C. Busso, “Jointly predicting arousal, valence and dominance with multi-task learning,” in INTERSPEECH 2017, Stockholm, Sweden, Aug. 2017.
Y. Li, J. Tao, B. Schuller, S. Shan, D. Jiang, and J. Jia,
“MEC 2016: The multimodal emotion recognition challenge of ccpr
Chinese Conference on Pattern Recognition 2016, pp. 667–678. Springer, 2016.
-  Niko Brummer, “Focal,” https://sites.google.com/site/nikobrummer/focal, 2017, Retrieved Aug 1st, 2017.
-  F. Eyben, M. Wöllmer, A. Graves, B. Schuller, E. Douglas-Cowie, and R. Cowie, “On-line emotion recognition in a 3-D activation-valence-time continuum using acoustic and linguistic cues,” Journal on Multimodal User Interfaces, vol. 3, no. 1-2, pp. 7–19, March 2010.
-  N. Dehak, P.J. Kenny, R. Dehak, P. Dumouchel, and P. Ouellet, “Front-end factor analysis for speaker verification,” IEEE Transactions on Audio, Speech, and Language Processing, vol. 19, no. 4, pp. 788–798, 2010.
-  R. Xia and Y. Liu, “Dbn-ivector framework for acoustic emotion recognition.,” in INTERSPEECH 2016, San Francisco,U.S.A, Sept. 2016, pp. 480–484.
-  G. Liu, Q. Qian, Z. Wang, Q. Zhao, T. Wang, H. Li, J. Xue, S. Zhu, R. Jin, and T. Zhao, “The opensesame NIST 2016 speaker recognition evaluation system,” in Interspeech 2017, Stockholm, Sweden, Sep. 2017, pp. 2854–2858.
-  V. Peddinti, D. Povey, and S. Khudanpur, “A time delay neural network architecture for efficient modeling of long temporal contexts.,” in INTERSPEECH 2015, Dresden, Germany, Sept. 2015, pp. 3214–3218.
-  H. Yu, C. Ho, Y. Juan, and C. Lin, “Libshorttext: A library for short-text classification and analysis,” Rapport interne, Department of Computer Science, National Taiwan University. Software available at http://www. csie. ntu. edu. tw/~ cjlin/libshorttext, 2013.
-  G. Parascandolo, H. Huttunen, and T. Virtanen, “Recurrent neural networks for polyphonic sound event detection in real life recordings,” arXiv preprint arXiv:1604.00861, 2016.
-  E. Cakir, G. Parascandolo, T. Heittola, H. Huttunen, and T. Virtanen, “Convolutional recurrent neural networks for polyphonic sound event detection,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 25, no. 6, pp. 1291–1303, 2017.
-  Dzmitry D. Bahdanau, J. Chorowski, D. Serdyuk, and Yoshua Y. Bengio, “End-to-end attention-based large vocabulary speech recognition,” in ICASSP 2016, Shanghai, China, Apr. 2016, IEEE, pp. 4945–4949.
-  F. Tao and G. Liu, “Advanced LSTM: A study about better time dependency modeling in emotion recognition,” in ICASSP 2018, Calgary, Canada, Apr. 2017.