| { | |
| "paper_id": "I17-1043", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T07:38:21.083848Z" | |
| }, | |
| "title": "Reusing Neural Speech Representations for Auditory Emotion Recognition", | |
| "authors": [ | |
| { | |
| "first": "Egor", | |
| "middle": [], | |
| "last": "Lakomkin", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Knowledge Technology University of Hamburg", | |
| "location": { | |
| "addrLine": "Vogt-Koelln Str. 30", | |
| "postCode": "22527", | |
| "settlement": "Hamburg", | |
| "country": "Germany" | |
| } | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Cornelius", | |
| "middle": [], | |
| "last": "Weber", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Knowledge Technology University of Hamburg", | |
| "location": { | |
| "addrLine": "Vogt-Koelln Str. 30", | |
| "postCode": "22527", | |
| "settlement": "Hamburg", | |
| "country": "Germany" | |
| } | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Sven", | |
| "middle": [], | |
| "last": "Magg", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Knowledge Technology University of Hamburg", | |
| "location": { | |
| "addrLine": "Vogt-Koelln Str. 30", | |
| "postCode": "22527", | |
| "settlement": "Hamburg", | |
| "country": "Germany" | |
| } | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Stefan", | |
| "middle": [], | |
| "last": "Wermter", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Knowledge Technology University of Hamburg", | |
| "location": { | |
| "addrLine": "Vogt-Koelln Str. 30", | |
| "postCode": "22527", | |
| "settlement": "Hamburg", | |
| "country": "Germany" | |
| } | |
| }, | |
| "email": "" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "Acoustic emotion recognition aims to categorize the affective state of the speaker and is still a difficult task for machine learning models. The difficulties come from the scarcity of training data, general subjectivity in emotion perception resulting in low annotator agreement, and the uncertainty about which features are the most relevant and robust ones for classification. In this paper, we will tackle the latter problem. Inspired by the recent success of transfer learning methods we propose a set of architectures which utilize neural representations inferred by training on large speech databases for the acoustic emotion recognition task. Our experiments on the IEMOCAP dataset show 10% relative improvements in the accuracy and F1-score over the baseline recurrent neural network which is trained endto-end for emotion recognition.", | |
| "pdf_parse": { | |
| "paper_id": "I17-1043", | |
| "_pdf_hash": "", | |
| "abstract": [ | |
| { | |
| "text": "Acoustic emotion recognition aims to categorize the affective state of the speaker and is still a difficult task for machine learning models. The difficulties come from the scarcity of training data, general subjectivity in emotion perception resulting in low annotator agreement, and the uncertainty about which features are the most relevant and robust ones for classification. In this paper, we will tackle the latter problem. Inspired by the recent success of transfer learning methods we propose a set of architectures which utilize neural representations inferred by training on large speech databases for the acoustic emotion recognition task. Our experiments on the IEMOCAP dataset show 10% relative improvements in the accuracy and F1-score over the baseline recurrent neural network which is trained endto-end for emotion recognition.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Abstract", | |
| "sec_num": null | |
| } | |
| ], | |
| "body_text": [ | |
| { | |
| "text": "Speech emotion recognition (SER) has received growing interest and attention in recent years. Being able to predict the affective state of a person gives valuable information which could improve dialog systems in human-computer interaction. To fully understand a current emotion expressed by a person, also knowledge of the context is required, like facial expressions, the semantics of a spoken text, gestures and body language, and cultural peculiarities. This makes it challenging even for people that have all this information to accurately predict the affective state. In this work, we are focusing solely on inferring the speaker's emotional state by analysing acoustic signals which are also the only source of information in situations when With recent advances in deep learning, which made it possible to train large end-to-end models for image classification (Simonyan and Zisserman, 2014) , speech recognition (Hannun et al., 2014) and natural language understanding (Sutskever et al., 2014) , the majority of the current work in the area of acoustic emotion recognition is neural network-based. Diverse neural architectures were investigated based on convolutional and recurrent neural networks (Fayek et al., 2017; Trigeorgis et al., 2016) . Alternatively, methods based on linear models, like SVM with careful feature engineering, still show competitive performance on the benchmark datasets (Schuller et al., , 2009 . Such methods were popular in computer vision until the AlexNet approach (Krizhevsky et al., 2012) made automatic feature learning more wide-spread. The low availability of annotated auditory emotion data is probably one of the main reasons for traditional methods being competitive. The appealing property of neural networks compared to SVM-like methods is their ability to identify automatically useful patterns in the data and to scale linearly with the number of training samples. These properties drive the research community to investigate different neural architectures.", | |
| "cite_spans": [ | |
| { | |
| "start": 869, | |
| "end": 899, | |
| "text": "(Simonyan and Zisserman, 2014)", | |
| "ref_id": "BIBREF20" | |
| }, | |
| { | |
| "start": 921, | |
| "end": 942, | |
| "text": "(Hannun et al., 2014)", | |
| "ref_id": "BIBREF11" | |
| }, | |
| { | |
| "start": 978, | |
| "end": 1002, | |
| "text": "(Sutskever et al., 2014)", | |
| "ref_id": "BIBREF21" | |
| }, | |
| { | |
| "start": 1207, | |
| "end": 1227, | |
| "text": "(Fayek et al., 2017;", | |
| "ref_id": "BIBREF7" | |
| }, | |
| { | |
| "start": 1228, | |
| "end": 1252, | |
| "text": "Trigeorgis et al., 2016)", | |
| "ref_id": "BIBREF22" | |
| }, | |
| { | |
| "start": 1406, | |
| "end": 1430, | |
| "text": "(Schuller et al., , 2009", | |
| "ref_id": "BIBREF19" | |
| }, | |
| { | |
| "start": 1505, | |
| "end": 1530, | |
| "text": "(Krizhevsky et al., 2012)", | |
| "ref_id": "BIBREF14" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "In this paper, we present a model that neither solely learns feature representations from scratch nor uses complex feature engineering but uses the features learned by a speech recognition network.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "Even though the automatic speech recognition (ASR) task is agnostic to the speaker's emotion and focuses only on the accuracy of the language transcription, low-level neural network layers trained for ASR might still extract useful information for the task of SER. Given that the size of the data suitable for training ASR systems is significantly larger than SER data, we can expect that trained ASR systems are more robust to speaker and condition variations. Recently, a method of transferring knowledge learned by the neural network from one task to another has been proposed (Rusu et al., 2016; Anderson et al., 2016) . This strategy also potentially prevents a neural network from overfitting to the smaller of the two datasets and could work as an additional regularizer. Moreover, in many applications, such as dialog systems, we would need to transcribe spoken text and identify its emotion jointly.", | |
| "cite_spans": [ | |
| { | |
| "start": 580, | |
| "end": 599, | |
| "text": "(Rusu et al., 2016;", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 600, | |
| "end": 622, | |
| "text": "Anderson et al., 2016)", | |
| "ref_id": "BIBREF1" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "In this paper, we evaluate several dual architectures which integrate representations of the ASR network: a fine-tuning and a progressive network. The fine-tuning architecture reuses features learnt by the recurrent layers of a speech recognition network and can use them directly for emotion classification by feeding them to a softmax classifier or can add additional hidden SER layers to tune ASR representations. Additionally, the ASR layers can be static for the whole training process or can be updated as well by allowing to backpropagate through them. The progressive architecture complements information from the ASR network with SER representations trained end-to-end. Therefore, in contrast to the fine-tuning model, a progressive network allows learning such low-level emotion-relevant features which the ASR network never learns since they are irrelevant to the speech recognition task. Our contribution in this paper is two-fold: 1) we propose several neural architectures allowing to model speech and emotion recognition jointly, and 2) we present a simple variant of a fine-tuning and a progressive network which improves the performance of the existing end-to-end models. ", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "The majority of recent research is aimed at searching for optimal neural architectures that learn emotion-specific features from little or not processed data. An autoencoder network (Ghosh et al., 2016) was demonstrated to learn to compress the speech frames before the emotion classification. An attention mechanism (Huang and Narayanan, 2016) was proposed to adjust weights for each of the speech frames depending on their importance. As there are many speech frames that are not relevant to an expressed emotion, such as silence, the attention mechanism allows focusing only on the significant part of the acoustic signal. Another approach probabilistically labeling each speech frame as emotional and non-emotional was proposed by (Chernykh et al., 2017) . A combination of convolutional and recurrent neural networks was demonstrated by (Trigeorgis et al., 2016 ) by training a model directly from the raw, unprocessed waveform, significantly outperforming manual feature engineering methods.", | |
| "cite_spans": [ | |
| { | |
| "start": 182, | |
| "end": 202, | |
| "text": "(Ghosh et al., 2016)", | |
| "ref_id": "BIBREF9" | |
| }, | |
| { | |
| "start": 317, | |
| "end": 344, | |
| "text": "(Huang and Narayanan, 2016)", | |
| "ref_id": "BIBREF12" | |
| }, | |
| { | |
| "start": 735, | |
| "end": 758, | |
| "text": "(Chernykh et al., 2017)", | |
| "ref_id": "BIBREF4" | |
| }, | |
| { | |
| "start": 842, | |
| "end": 866, | |
| "text": "(Trigeorgis et al., 2016", | |
| "ref_id": "BIBREF22" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Related work", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "One of the first examples of the knowledge transfer among neural networks was demonstrated by (Bengio, 2012) and (Yosinski et al., 2014) . Eventually, fine-tuning became the de-facto standard for computer vision tasks that have a small number of annotated samples, leveraging the ability of trained convolutional filters to be applicable to different tasks. Our work is mainly inspired by recently introduced architectures with an ability to transfer knowledge between recurrent neural networks in a domain different from computer vision (Rusu et al., 2016; Anderson et al., 2016) .", | |
| "cite_spans": [ | |
| { | |
| "start": 94, | |
| "end": 108, | |
| "text": "(Bengio, 2012)", | |
| "ref_id": "BIBREF2" | |
| }, | |
| { | |
| "start": 113, | |
| "end": 136, | |
| "text": "(Yosinski et al., 2014)", | |
| "ref_id": "BIBREF23" | |
| }, | |
| { | |
| "start": 538, | |
| "end": 557, | |
| "text": "(Rusu et al., 2016;", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 558, | |
| "end": 580, | |
| "text": "Anderson et al., 2016)", | |
| "ref_id": "BIBREF1" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Related work", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "Previously, to our knowledge, there was only one attempt to analyze the correlation between automatic speech and emotion recognition (Fayek et al., 2016) . This approach showed the possibility of knowledge transfer from the convolutional neural acoustic model trained on the TIMIT corpus (Garofolo et al., 1993) for the emotion recog-nition task. The authors proposed several variants of fine-tuning. They reported a significant drop in the performance by using the ASR network as a feature extractor and training only the output softmax layer, compared to an end-to-end convolution neural network model. Gradual improvements were observed by allowing more ASR layers to be updated during back-propagation but, overall, using ASR for feature extraction affected the performance negatively.", | |
| "cite_spans": [ | |
| { | |
| "start": 133, | |
| "end": 153, | |
| "text": "(Fayek et al., 2016)", | |
| "ref_id": "BIBREF6" | |
| }, | |
| { | |
| "start": 288, | |
| "end": 311, | |
| "text": "(Garofolo et al., 1993)", | |
| "ref_id": "BIBREF8" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Related work", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "3 Models and experiment setup", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Related work", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "We introduce two models that use a pre-trained ASR network for acoustic emotion recognition. The first one is the fine-tuning model which only takes representations of the ASR network and learns how to combine them to predict an emotion category. We compare two variants of tuning ASR representations: simply feeding them into a softmax classifier or adding a new Gated Recurrent Units (GRU) layer trained on top of the ASR features. The second is the progressive network which allows us to train a neural network branch parallel to the ASR network which can capture additional emotion-specific information. We present all SER models used in this work in Figure 3 .", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 655, | |
| "end": 663, | |
| "text": "Figure 3", | |
| "ref_id": "FIGREF2" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Models", | |
| "sec_num": "3.1" | |
| }, | |
| { | |
| "text": "Our ASR model (see Figure 2 ) is a combination of convolutional and recurrent layers inspired by the DeepSpeech (Hannun et al., 2014) architecture for speech recognition. Our model contains two convolutional layers for feature extraction from power FFT spectrograms, followed by five recurrent bi-directional GRU layers with the softmax layer on top, predicting the character distribution for each speech frame (ASR network in all our experiments, left branch of the network in the Figures 3b, 3c and 3d ). The ASR network is trained on pairs of utterances and the corresponding transcribed texts (see 3.2.2 \"Speech data\" section for details). Connectionist Temporal classification (CTC) loss (Graves et al., 2006) was used as a metric to measure how good the alignment produced by the network is compared to the ground truth transcription. Power spectrograms were extracted using a Hamming window of 20ms width and 10 ms stride, resulting in 161 features for each speech frame. We trained the ASR network with Stochastic Gradient Descent with a learning rate of 0.0003 divided by 1.1 after every epoch until the character error rate stopped improving on the validation set (resulting in 35 epochs overall). In all our experiments we keep the ASR network static by freezing its weights during training for SER.", | |
| "cite_spans": [ | |
| { | |
| "start": 112, | |
| "end": 133, | |
| "text": "(Hannun et al., 2014)", | |
| "ref_id": "BIBREF11" | |
| }, | |
| { | |
| "start": 693, | |
| "end": 714, | |
| "text": "(Graves et al., 2006)", | |
| "ref_id": "BIBREF10" | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 19, | |
| "end": 27, | |
| "text": "Figure 2", | |
| "ref_id": "FIGREF1" | |
| }, | |
| { | |
| "start": 482, | |
| "end": 503, | |
| "text": "Figures 3b, 3c and 3d", | |
| "ref_id": "FIGREF2" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "ASR model", | |
| "sec_num": "3.1.1" | |
| }, | |
| { | |
| "text": "As a baseline (see Figure 3a ), we used a twolayer bi-directional GRU neural network. Utterances were represented by averaging hidden vector representations obtained on the second GRU layer and fed to a softmax layer for emotion classification. Dropout with the probability of 0.25 was applied to the utterance representation during training to prevent overfitting. We evaluate this architecture as a baseline as it was proven to yield strong results on the acoustic emotion recognition task (Huang and Narayanan, 2016) . As there are significantly less emotion-annotated samples available the SER-specific network is limited to two layers compared to the ASR network.", | |
| "cite_spans": [ | |
| { | |
| "start": 492, | |
| "end": 519, | |
| "text": "(Huang and Narayanan, 2016)", | |
| "ref_id": "BIBREF12" | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 19, | |
| "end": 28, | |
| "text": "Figure 3a", | |
| "ref_id": "FIGREF2" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Baseline", | |
| "sec_num": "3.1.2" | |
| }, | |
| { | |
| "text": "We propose several variants of reusing speech representations: 1) By averaging hidden memory representations of the layer number x of the ASR network (Fine-tuning MP-x later in the text where MP stands for Mean Pooling), we train only the output softmax layer to predict an emotion class.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Fine-tuning model", | |
| "sec_num": "3.1.3" | |
| }, | |
| { | |
| "text": "(2) We feed hidden memory representations as input to a new GRU network (emotion-specific) initialized randomly and trained for emotion classification (Fine-tuning RNN-x). The intuition is that bottom layers of the ASR network can be used as feature extractors and the top level GRU can combine them to predict the emotion class. Similar to the Fine-tuning MP-x setup we average representations of the newly attached GRU layer and feed them to the classifier. In both experiments, dropout with the rate of 0.25 was applied to averaged representations. Figure 3b shows the Finetuning MP-1 model pooling ASR representation of the first ASR layer, and Figure 3c shows the Finetuning RNN-1 setup.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 552, | |
| "end": 561, | |
| "text": "Figure 3b", | |
| "ref_id": "FIGREF2" | |
| }, | |
| { | |
| "start": 649, | |
| "end": 658, | |
| "text": "Figure 3c", | |
| "ref_id": "FIGREF2" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Fine-tuning model", | |
| "sec_num": "3.1.3" | |
| }, | |
| { | |
| "text": "The progressive network combines fine-tuning with end-to-end training. The scheme of the model is presented in Figure 3d . It contains two branches: first, a speech recognition branch (left in Figure 3d and identical to Figure 2 ) which is static and not updated, and second, an emotion recogni- tion branch (right), with the same architecture as the baseline model which we initialized randomly and trained from scratch. We feed the same features to the emotion recognition branch as to the baseline model for a fair comparison. Theoretically, a network of such type can learn task specific features while incorporating knowledge already utilized in the ASR network if it contributes positively to the prediction.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 111, | |
| "end": 120, | |
| "text": "Figure 3d", | |
| "ref_id": "FIGREF2" | |
| }, | |
| { | |
| "start": 193, | |
| "end": 202, | |
| "text": "Figure 3d", | |
| "ref_id": "FIGREF2" | |
| }, | |
| { | |
| "start": 220, | |
| "end": 228, | |
| "text": "Figure 2", | |
| "ref_id": "FIGREF1" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Progressive neural network", | |
| "sec_num": "3.1.4" | |
| }, | |
| { | |
| "text": "The Interactive Emotional Dyadic Motion Capture dataset IEMOCAP (Busso et al., 2008) contains five recorded sessions of conversations between two actors, one from each gender. The total amount of data is 12 hours of audio-visual information from ten speakers annotated with categorical emotion labels (Anger, Happiness, Sadness, Neutral, Surprise, Fear, Frustration and Excited), and dimensional labels (values of the activation and valence from 1 to 5). Similarly as in previous work (Huang and Narayanan, 2016) , we merged the Excited class with Happiness. We performed several data filtering steps: we kept samples where at least two annotators agreed on the emotion la-bel, discarded samples where an utterance was annotated with 3 different emotions and used samples annotated with neutral, angry, happy and sad, resulting in 6,416 samples (1,104 of Anger, 2,496 of Happiness, 1,752 of Neutral and 1,064 of Sadness). We use 4 out of 5 sessions for training and the remaining one for validation and testing (as there are two speakers in the session, one was used for validation and the other for testing).", | |
| "cite_spans": [ | |
| { | |
| "start": 64, | |
| "end": 84, | |
| "text": "(Busso et al., 2008)", | |
| "ref_id": "BIBREF3" | |
| }, | |
| { | |
| "start": 485, | |
| "end": 512, | |
| "text": "(Huang and Narayanan, 2016)", | |
| "ref_id": "BIBREF12" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Emotion data", | |
| "sec_num": "3.2.1" | |
| }, | |
| { | |
| "text": "We concatenated three datasets to train the ASR model: LibriSpeech, TED-LIUM v2, and Vox-Forge. LibriSpeech (Panayotov et al., 2015) contains around 1,000 hours of English-read speech from audiobooks. TED-LIUM v2 (Rousseau et al., 2014 ) is a dataset composed of transcribed TED talks, containing 200 hours of speech and 1495 speakers. VoxForge is an open-source collection of transcribed recordings collected using crowdsourcing. We downloaded all English recordings 1 , which is around 100 hours of speech. Overall, 384,547 utterances containing 1,300 hours of speech from more than 3,000 speakers were used Figure 4 : Representations of the IEMOCAP utterances generated by the Fine-tuning MP-x and Finetuning RNN-x networks (x stands for the ASR layer number of which the representation is used) projected into 2-dimensional space using the t-SNE technique. Top: all four classes, bottom: only Sadness and Anger. Color mapping: Anger -red, Sadness -blue, Neutral -cyan, Happiness -green. We can observe that Fine-tuning MP-2 and MP-3 networks can separate Anger and Sadness classes even though these representations are directly computed from the ASR network without any emotion-specific training. Fine-tuning RNN networks benefit from being trained directly for emotion recognition and form visually distinguishable clusters.", | |
| "cite_spans": [ | |
| { | |
| "start": 108, | |
| "end": 132, | |
| "text": "(Panayotov et al., 2015)", | |
| "ref_id": "BIBREF15" | |
| }, | |
| { | |
| "start": 213, | |
| "end": 235, | |
| "text": "(Rousseau et al., 2014", | |
| "ref_id": "BIBREF16" | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 610, | |
| "end": 618, | |
| "text": "Figure 4", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Speech data", | |
| "sec_num": "3.2.2" | |
| }, | |
| { | |
| "text": "to train the ASR model. We conducted no preprocessing other than the conversion of recordings to WAV format with single channel 32-bit format and a sampling rate of 16,000. Utterances longer than 15 seconds were filtered out due to GPU memory constraints.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Speech data", | |
| "sec_num": "3.2.2" | |
| }, | |
| { | |
| "text": "The ASR network was trained on power spectrograms with filter banks computed over windows of 20ms width and 10ms stride. For the progressive network, we used the same features as for the ASR branch, and for the SER-specific branch, we used 13 MFCC coefficients and their deltas extracted over windows of 20ms width and 10ms stride. We have extracted pitch values smoothed with moving average with a window size of 15 using the OpenSMILE toolkit . The reason for the choice of high-level features like MFCC and pitch for the SER-branch was the limited size of emotion annotated data, as learning efficient emotion representation from low-level features like raw waveform or power-spectrograms might be difficult on the dataset of a size of IEMO-CAP. In addition, such feature set showed state-ofthe-art results (Huang and Narayanan, 2016) . In both cases, we normalized each feature by subtracting the mean and dividing by standard deviation per utterance.", | |
| "cite_spans": [ | |
| { | |
| "start": 810, | |
| "end": 837, | |
| "text": "(Huang and Narayanan, 2016)", | |
| "ref_id": "BIBREF12" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Extracted features", | |
| "sec_num": "3.3" | |
| }, | |
| { | |
| "text": "The Adam optimizer (Kingma and Ba, 2014) was used in all experiments with a learning rate of 0.0001, clipping the norm of the gradient at the level of 15 with a batch size of 64. During the training, we applied learning rate annealing if the results on the validation set did not improve for two epochs and stopped it when the learning rate reaches the value of 1e-6. We applied the Sorta-Grad algorithm (Amodei et al., 2015) during the first epoch by sorting utterances by the duration (Hannun et al., 2014) . We performed data augmentation during the training phase by randomly changing tempo and gain of the utterance within the range of 0.85 to 1.15 and -3 to +6 dB respectively. The model with the lowest cross-entropy loss on the validation set was picked to evaluate the test set performance. Table 1 : Utterance-level emotion 4-way classification performance (unweighted and weighted accuracy and f1-score). Several variants of fine-tuning and progressive networks are evaluated: using first, second or third ASR layers as input for Fine-tuning MP, Fine-tuning RNN, and progressive networks. Table 1 summarizes the results obtained from ASR-SER transfer learning. We evaluate several baseline models by varying the number of GRU units in a network, and three variants for Finetuning MP-x, Fine-tuning RNN-x and Progressive net-x by utilizing representations of layer x of the ASR network. We report weighted and unweighted accuracy and f1-score to reflect imbalanced classes. These metrics were averaged over ten runs of a ten-fold leave-one-speaker-out crossvalidation to monitor an effect of random initialization of a neural network. Also, our results reveal the difficulty of separating Anger and Happiness classes, and Neutral and Happiness (see Figure 5 ). Our best Fine-tuning MP-3 model achieved 55% unweighted and 56% weighted accuracy, which significantly outperforms the baseline (p-value \u2264 0.03) end-to-end 2-layer GRU neural network similar to (Huang and Narayanan, 2016) and (Ghosh et al., 2016) . The fine-tuning model has around 30 times less trainable parameters (as only the softmax layer is trained) and achieves significantly better performance than the baseline. These results show that putting an additional GRU layer on top of ASR representations affects the performance positively and shows significantly better results than the baseline (p-value \u2264 0.0007). Results prove our hypothesis that intermediate features extracted by the ASR network contain useful information for emotion classification. The progressive network consistently outperforms baseline end-to-end models, reaching 58% unweighted and 61% weighted accuracy. In all variants, the addition of the second recurrent layer representations of the ASR's network contributes positively to the performance compared to the baseline. Our results support the hypothesis that the progressive architecture of the network allows to combine the ASR low-level representations with the SER-specific ones and achieve the best accuracy result.", | |
| "cite_spans": [ | |
| { | |
| "start": 404, | |
| "end": 425, | |
| "text": "(Amodei et al., 2015)", | |
| "ref_id": "BIBREF0" | |
| }, | |
| { | |
| "start": 487, | |
| "end": 508, | |
| "text": "(Hannun et al., 2014)", | |
| "ref_id": "BIBREF11" | |
| }, | |
| { | |
| "start": 1965, | |
| "end": 1992, | |
| "text": "(Huang and Narayanan, 2016)", | |
| "ref_id": "BIBREF12" | |
| }, | |
| { | |
| "start": 1997, | |
| "end": 2017, | |
| "text": "(Ghosh et al., 2016)", | |
| "ref_id": "BIBREF9" | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 800, | |
| "end": 807, | |
| "text": "Table 1", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 1100, | |
| "end": 1107, | |
| "text": "Table 1", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 1759, | |
| "end": 1767, | |
| "text": "Figure 5", | |
| "ref_id": "FIGREF3" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Training", | |
| "sec_num": "3.3.1" | |
| }, | |
| { | |
| "text": "In addition to the quantitative results, we tried to analyze the reason of such effectiveness of the ASR representations, by visualizing the representations of the utterances by Fine-tuning MP-x and Fine-tuning RNN-x networks (see Figure 4) . We observe that a prior ASR-trained network can separate Sadness and Anger samples even by pooling representations of the first ASR layer. On the 3rd layer, Anger, Sadness and Happiness form visually distinguishable clusters which could explain the surprising effectiveness of Fine-tuning MP-2/3 models. Fine-tuning RNN-x networks can separate four classes better due to an additional trained GRU network on top of the ASR representations. Also, we found that activations of some neurons in the ASR network correlate significantly with the well-known prosodic features like loudness. Figure 6 shows the activation of the neuron number 840 of the second GRU layer of the ASR network and the loudness value of the speech frame for two audio files. We found that, on average, the Pearson's correlation between loudness and activation of the 840th neuron calculated on the IEMOCAP dataset is greater than 0.64 which is an indicator that the ASR network is capable of learning prosodic features which is useful for emotion classification.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 231, | |
| "end": 240, | |
| "text": "Figure 4)", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 827, | |
| "end": 835, | |
| "text": "Figure 6", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Results", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "In this paper, various neural architectures were proposed utilizing speech recognition representations. Fine-tuning provides an ability to use the ASR network for the emotion recognition task quickly. A progressive network allows to combine speech and emotion representations and train them in parallel. Our experimental results confirm that trained speech representations, even though expected to be agnostic to a speaker's emotion, Figure 6 : Activations of the neuron number 840 of the second GRU layer of the ASR model (blue) and loudness of two utterances selected randomly from IEMOCAP dataset (orange).", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 434, | |
| "end": 442, | |
| "text": "Figure 6", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Discussion and conclusion", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "contain useful information for affective state predictions.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Discussion and conclusion", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "A possible future research direction would be to investigate the influence of linguistic knowledge on the speech representations and how it affects the system performance. Additionally, the ASR system can be fine-tuned in parallel with the emotion branch by updating the layers of the ASR network. Potentially, this could help the system to adapt better to the particular speakers and their emotion expression style. Furthermore, analyzing linguistic information of the spoken text produced by the ASR network could possibly alleviate the difficulty of separating Anger and Happiness classes, and Neutral and Happiness.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Discussion and conclusion", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "http://www.repository.voxforge1. org/downloads/SpeechCorpus/Trunk/Audio/ Original/48kHz_16bit/", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [ | |
| { | |
| "text": "This project has received funding from the European Union's Horizon 2020 research and innovation programme under the Marie Sklodowska-Curie grant agreement No 642667 (SECURE) and partial support from the German Research Foundation DFG under project CML (TRR 169).", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Acknowledgments", | |
| "sec_num": null | |
| } | |
| ], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "Deep speech 2: End-to-end speech recognition in English and Mandarin", | |
| "authors": [ | |
| { | |
| "first": "Dario", | |
| "middle": [], | |
| "last": "Amodei", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Rishita", | |
| "middle": [], | |
| "last": "Anubhai", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Eric", | |
| "middle": [], | |
| "last": "Battenberg", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Carl", | |
| "middle": [], | |
| "last": "Case", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Jared", | |
| "middle": [], | |
| "last": "Casper", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2015, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Dario Amodei, Rishita Anubhai, Eric Battenberg, Carl Case, Jared Casper, and et al. 2015. Deep speech 2: End-to-end speech recognition in English and Man- darin. CoRR, abs/1512.02595.", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "Beyond Fine Tuning: A Modular Approach to Learning on Small Data", | |
| "authors": [ | |
| { | |
| "first": "Ark", | |
| "middle": [], | |
| "last": "Anderson", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Kyle", | |
| "middle": [], | |
| "last": "Shaffer", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Artem", | |
| "middle": [], | |
| "last": "Yankov", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Court", | |
| "middle": [], | |
| "last": "Corley", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Nathan", | |
| "middle": [], | |
| "last": "Hodas", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2016, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Ark Anderson, Kyle Shaffer, Artem Yankov, Court Corley, and Nathan Hodas. 2016. Beyond Fine Tun- ing: A Modular Approach to Learning on Small Data. CoRR, abs/1611.01714.", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "Deep learning of representations for unsupervised and transfer learning", | |
| "authors": [ | |
| { | |
| "first": "Yoshua", | |
| "middle": [], | |
| "last": "Bengio", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2012, | |
| "venue": "ICML Unsupervised and Transfer Learning", | |
| "volume": "27", | |
| "issue": "", | |
| "pages": "17--36", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Yoshua Bengio. 2012. Deep learning of representa- tions for unsupervised and transfer learning. ICML Unsupervised and Transfer Learning, 27:17-36.", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "IEMOCAP: interactive emotional dyadic motion capture database. Language Resources and Evaluation", | |
| "authors": [ | |
| { | |
| "first": "Carlos", | |
| "middle": [], | |
| "last": "Busso", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Murtaza", | |
| "middle": [], | |
| "last": "Bulut", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Chi-Chun", | |
| "middle": [], | |
| "last": "Lee", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Abe", | |
| "middle": [], | |
| "last": "Kazemzadeh", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Emily", | |
| "middle": [], | |
| "last": "Mower", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Samuel", | |
| "middle": [], | |
| "last": "Kim", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Jeannette", | |
| "middle": [ | |
| "N" | |
| ], | |
| "last": "Chang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Sungbok", | |
| "middle": [], | |
| "last": "Lee", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Shrikanth", | |
| "middle": [ | |
| "S" | |
| ], | |
| "last": "Narayanan", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2008, | |
| "venue": "", | |
| "volume": "42", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Carlos Busso, Murtaza Bulut, Chi-Chun Lee, Abe Kazemzadeh, Emily Mower, Samuel Kim, Jean- nette N. Chang, Sungbok Lee, and Shrikanth S. Narayanan. 2008. IEMOCAP: interactive emotional dyadic motion capture database. Language Re- sources and Evaluation, 42(4):335.", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "Emotion recognition from speech with recurrent neural networks", | |
| "authors": [ | |
| { | |
| "first": "Vladimir", | |
| "middle": [], | |
| "last": "Chernykh", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Grigoriy", | |
| "middle": [], | |
| "last": "Sterling", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Pavel", | |
| "middle": [], | |
| "last": "Prihodko", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2017, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Vladimir Chernykh, Grigoriy Sterling, and Pavel Pri- hodko. 2017. Emotion recognition from speech with recurrent neural networks. CoRR, abs/1701.08071.", | |
| "links": null | |
| }, | |
| "BIBREF5": { | |
| "ref_id": "b5", | |
| "title": "Recent developments in opensmile, the munich open-source multimedia feature extractor", | |
| "authors": [ | |
| { | |
| "first": "Florian", | |
| "middle": [], | |
| "last": "Eyben", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Felix", | |
| "middle": [], | |
| "last": "Weninger", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Florian", | |
| "middle": [], | |
| "last": "Gross", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Bj\u00f6rn", | |
| "middle": [], | |
| "last": "Schuller", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "Proceedings of the 21st ACM International Conference on Multimedia", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "835--838", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Florian Eyben, Felix Weninger, Florian Gross, and Bj\u00f6rn Schuller. 2013. Recent developments in opensmile, the munich open-source multimedia fea- ture extractor. In Proceedings of the 21st ACM Inter- national Conference on Multimedia, pages 835-838. ACM.", | |
| "links": null | |
| }, | |
| "BIBREF6": { | |
| "ref_id": "b6", | |
| "title": "On the Correlation and Transferability of Features Between Automatic Speech Recognition and Speech Emotion Recognition", | |
| "authors": [ | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Haytham", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Margaret", | |
| "middle": [], | |
| "last": "Fayek", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Lawrence", | |
| "middle": [], | |
| "last": "Lech", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Cavedon", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2016, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "3618--3622", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Haytham M. Fayek, Margaret Lech, and Lawrence Cavedon. 2016. On the Correlation and Trans- ferability of Features Between Automatic Speech Recognition and Speech Emotion Recognition. pages 3618-3622.", | |
| "links": null | |
| }, | |
| "BIBREF7": { | |
| "ref_id": "b7", | |
| "title": "Evaluating deep learning architectures for Speech Emotion Recognition", | |
| "authors": [ | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Haytham", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Margaret", | |
| "middle": [], | |
| "last": "Fayek", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Lawrence", | |
| "middle": [], | |
| "last": "Lech", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Cavedon", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2017, | |
| "venue": "Neural Networks", | |
| "volume": "92", | |
| "issue": "", | |
| "pages": "60--68", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Haytham M. Fayek, Margaret Lech, and Lawrence Cavedon. 2017. Evaluating deep learning architec- tures for Speech Emotion Recognition. Neural Net- works, vol. 92:60-68.", | |
| "links": null | |
| }, | |
| "BIBREF8": { | |
| "ref_id": "b8", | |
| "title": "DARPA TIMIT acousticphonetic continuous speech corpus CD-ROM. NIST speech disc 1-1.1. NASA STI/Recon Technical Report N", | |
| "authors": [ | |
| { | |
| "first": "J", | |
| "middle": [ | |
| "S" | |
| ], | |
| "last": "Garofolo", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "L", | |
| "middle": [ | |
| "F" | |
| ], | |
| "last": "Lamel", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "W", | |
| "middle": [ | |
| "M" | |
| ], | |
| "last": "Fisher", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [ | |
| "G" | |
| ], | |
| "last": "Fiscus", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "D", | |
| "middle": [ | |
| "S" | |
| ], | |
| "last": "Pallett", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1993, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "J. S. Garofolo, L. F. Lamel, W. M. Fisher, J. G. Fiscus, and D. S. Pallett. 1993. DARPA TIMIT acoustic- phonetic continuous speech corpus CD-ROM. NIST speech disc 1-1.1. NASA STI/Recon Technical Re- port N, 93.", | |
| "links": null | |
| }, | |
| "BIBREF9": { | |
| "ref_id": "b9", | |
| "title": "Representation Learning for Speech Emotion Recognition", | |
| "authors": [ | |
| { | |
| "first": "Sayan", | |
| "middle": [], | |
| "last": "Ghosh", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Eugene", | |
| "middle": [], | |
| "last": "Laksana", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Louis-Philippe", | |
| "middle": [], | |
| "last": "Morency", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Stefan", | |
| "middle": [], | |
| "last": "Scherer", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2016, | |
| "venue": "INTERSPEECH", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "3603--3607", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Sayan Ghosh, Eugene Laksana, Louis-Philippe Morency, and Stefan Scherer. 2016. Representa- tion Learning for Speech Emotion Recognition. INTERSPEECH, pages 3603-3607.", | |
| "links": null | |
| }, | |
| "BIBREF10": { | |
| "ref_id": "b10", | |
| "title": "Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks", | |
| "authors": [ | |
| { | |
| "first": "Alex", | |
| "middle": [], | |
| "last": "Graves", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Santiago", | |
| "middle": [], | |
| "last": "Fernandez", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Faustino", | |
| "middle": [], | |
| "last": "Gomez", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Jrgen", | |
| "middle": [], | |
| "last": "Schmidhuber", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2006, | |
| "venue": "Proceedings of the 23rd International Conference on Machine Learning", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "369--376", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Alex Graves, Santiago Fernandez, Faustino Gomez, and Jrgen Schmidhuber. 2006. Connectionist temporal classification: labelling unsegmented se- quence data with recurrent neural networks. In Pro- ceedings of the 23rd International Conference on Machine Learning, pages 369-376. ACM.", | |
| "links": null | |
| }, | |
| "BIBREF11": { | |
| "ref_id": "b11", | |
| "title": "Deep Speech: Scaling up endto-end speech recognition", | |
| "authors": [ | |
| { | |
| "first": "Awni", | |
| "middle": [], | |
| "last": "Hannun", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Carl", | |
| "middle": [], | |
| "last": "Case", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Jared", | |
| "middle": [], | |
| "last": "Casper", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Bryan", | |
| "middle": [], | |
| "last": "Catanzaro", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2014, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Awni Hannun, Carl Case, Jared Casper, Bryan Catan- zaro, and et al. 2014. Deep Speech: Scaling up end- to-end speech recognition. CoRR, abs/1412.5567.", | |
| "links": null | |
| }, | |
| "BIBREF12": { | |
| "ref_id": "b12", | |
| "title": "Attention Assisted Discovery of Sub-Utterance Structure in Speech Emotion Recognition", | |
| "authors": [ | |
| { | |
| "first": "Che-Wei", | |
| "middle": [], | |
| "last": "Huang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Shrikanth", | |
| "middle": [ | |
| "S" | |
| ], | |
| "last": "Narayanan", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2016, | |
| "venue": "Proceedings of Interspeech", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "1387--1391", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Che-Wei Huang and Shrikanth S. Narayanan. 2016. Attention Assisted Discovery of Sub-Utterance Structure in Speech Emotion Recognition. In Pro- ceedings of Interspeech, pages 1387-1391.", | |
| "links": null | |
| }, | |
| "BIBREF13": { | |
| "ref_id": "b13", | |
| "title": "Adam: A Method for Stochastic Optimization", | |
| "authors": [ | |
| { | |
| "first": "P", | |
| "middle": [], | |
| "last": "Diederik", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Jimmy", | |
| "middle": [], | |
| "last": "Kingma", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Ba", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2014, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": { | |
| "arXiv": [ | |
| "arXiv:1412.6980[cs].ArXiv:1412.6980" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Diederik P. Kingma and Jimmy Ba. 2014. Adam: A Method for Stochastic Optimization. arXiv:1412.6980 [cs]. ArXiv: 1412.6980.", | |
| "links": null | |
| }, | |
| "BIBREF14": { | |
| "ref_id": "b14", | |
| "title": "ImageNet Classification with Deep Convolutional Neural Networks", | |
| "authors": [ | |
| { | |
| "first": "Alex", | |
| "middle": [], | |
| "last": "Krizhevsky", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Ilya", | |
| "middle": [], | |
| "last": "Sutskever", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Geoffrey", | |
| "middle": [ | |
| "E" | |
| ], | |
| "last": "Hinton", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2012, | |
| "venue": "Advances in Neural Information Processing Systems", | |
| "volume": "25", | |
| "issue": "", | |
| "pages": "1097--1105", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hin- ton. 2012. ImageNet Classification with Deep Con- volutional Neural Networks. In F. Pereira, C. J. C. Burges, L. Bottou, and K. Q. Weinberger, editors, Advances in Neural Information Processing Systems 25, pages 1097-1105. Curran Associates, Inc.", | |
| "links": null | |
| }, | |
| "BIBREF15": { | |
| "ref_id": "b15", | |
| "title": "Librispeech: an ASR corpus based on public domain audio books", | |
| "authors": [ | |
| { | |
| "first": "Vassil", | |
| "middle": [], | |
| "last": "Panayotov", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Guoguo", | |
| "middle": [], | |
| "last": "Chen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Daniel", | |
| "middle": [], | |
| "last": "Povey", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Sanjeev", | |
| "middle": [], | |
| "last": "Khudanpur", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2015, | |
| "venue": "Acoustics, Speech and Signal Processing (ICASSP), 2015 IEEE International Conference on", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "5206--5210", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. 2015. Librispeech: an ASR corpus based on public domain audio books. In Acoustics, Speech and Signal Processing (ICASSP), 2015 IEEE International Conference on, pages 5206-5210. IEEE.", | |
| "links": null | |
| }, | |
| "BIBREF16": { | |
| "ref_id": "b16", | |
| "title": "Enhancing the ted-lium corpus with selected data for language modeling and more ted talks", | |
| "authors": [ | |
| { | |
| "first": "Anthony", | |
| "middle": [], | |
| "last": "Rousseau", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Paul", | |
| "middle": [], | |
| "last": "Del\u00e9glise", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Yannick", | |
| "middle": [], | |
| "last": "Est\u00e8ve", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2014, | |
| "venue": "LREC", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "3935--3939", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Anthony Rousseau, Paul Del\u00e9glise, and Yannick Est\u00e8ve. 2014. Enhancing the ted-lium corpus with selected data for language modeling and more ted talks. In LREC, pages 3935-3939.", | |
| "links": null | |
| }, | |
| "BIBREF18": { | |
| "ref_id": "b18", | |
| "title": "The interspeech 2013 computational paralinguistics challenge: social signals", | |
| "authors": [ | |
| { | |
| "first": "Bj\u00f6rn", | |
| "middle": [], | |
| "last": "Schuller", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Stefan", | |
| "middle": [], | |
| "last": "Steidl", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Anton", | |
| "middle": [], | |
| "last": "Batliner", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Alessandro", | |
| "middle": [], | |
| "last": "Vinciarelli", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Klaus", | |
| "middle": [], | |
| "last": "Scherer", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Fabien", | |
| "middle": [], | |
| "last": "Ringeval", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Mohamed", | |
| "middle": [], | |
| "last": "Chetouani", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Felix", | |
| "middle": [], | |
| "last": "Weninger", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Florian", | |
| "middle": [], | |
| "last": "Eyben", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Erik", | |
| "middle": [], | |
| "last": "Marchi", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Bj\u00f6rn Schuller, Stefan Steidl, Anton Batliner, Alessan- dro Vinciarelli, Klaus Scherer, Fabien Ringeval, Mo- hamed Chetouani, Felix Weninger, Florian Eyben, Erik Marchi, et al. 2013. The interspeech 2013 com- putational paralinguistics challenge: social signals, conflict, emotion, autism.", | |
| "links": null | |
| }, | |
| "BIBREF19": { | |
| "ref_id": "b19", | |
| "title": "Acoustic emotion recognition: A benchmark comparison of performances", | |
| "authors": [ | |
| { | |
| "first": "Bj\u00f6rn", | |
| "middle": [], | |
| "last": "Schuller", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Bogdan", | |
| "middle": [], | |
| "last": "Vlasenko", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Florian", | |
| "middle": [], | |
| "last": "Eyben", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Gerhard", | |
| "middle": [], | |
| "last": "Rigoll", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Andreas", | |
| "middle": [], | |
| "last": "Wendemuth", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2009, | |
| "venue": "Automatic Speech Recognition & Understanding", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "552--557", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Bj\u00f6rn Schuller, Bogdan Vlasenko, Florian Eyben, Gerhard Rigoll, and Andreas Wendemuth. 2009. Acoustic emotion recognition: A benchmark com- parison of performances. In Automatic Speech Recognition & Understanding, pages 552-557. IEEE.", | |
| "links": null | |
| }, | |
| "BIBREF20": { | |
| "ref_id": "b20", | |
| "title": "Very deep convolutional networks for large-scale image recognition", | |
| "authors": [ | |
| { | |
| "first": "Karen", | |
| "middle": [], | |
| "last": "Simonyan", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Andrew", | |
| "middle": [], | |
| "last": "Zisserman", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2014, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Karen Simonyan and Andrew Zisserman. 2014. Very deep convolutional networks for large-scale image recognition. CoRR, abs/1409.1556.", | |
| "links": null | |
| }, | |
| "BIBREF21": { | |
| "ref_id": "b21", | |
| "title": "Sequence to sequence learning with neural networks", | |
| "authors": [ | |
| { | |
| "first": "Ilya", | |
| "middle": [], | |
| "last": "Sutskever", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Oriol", | |
| "middle": [], | |
| "last": "Vinyals", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "V", | |
| "middle": [], | |
| "last": "Quoc", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Le", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2014, | |
| "venue": "Advances in neural information processing systems", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "3104--3112", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. 2014. Sequence to sequence learning with neural net- works. In Advances in neural information process- ing systems, pages 3104-3112.", | |
| "links": null | |
| }, | |
| "BIBREF22": { | |
| "ref_id": "b22", | |
| "title": "Adieu features? End-to-end speech emotion recognition using a deep convolutional recurrent network", | |
| "authors": [ | |
| { | |
| "first": "George", | |
| "middle": [], | |
| "last": "Trigeorgis", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Fabien", | |
| "middle": [], | |
| "last": "Ringeval", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Raymond", | |
| "middle": [], | |
| "last": "Brueckner", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Erik", | |
| "middle": [], | |
| "last": "Marchi", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Mihalis", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Bjrn", | |
| "middle": [], | |
| "last": "Nicolaou", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Stefanos", | |
| "middle": [], | |
| "last": "Schuller", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Zafeiriou", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2016, | |
| "venue": "Acoustics, Speech and Signal Processing", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "5200--5204", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "George Trigeorgis, Fabien Ringeval, Raymond Brueck- ner, Erik Marchi, Mihalis A. Nicolaou, Bjrn Schuller, and Stefanos Zafeiriou. 2016. Adieu fea- tures? End-to-end speech emotion recognition using a deep convolutional recurrent network. In Acous- tics, Speech and Signal Processing (ICASSP), 2016 IEEE International Conference on, pages 5200- 5204. IEEE.", | |
| "links": null | |
| }, | |
| "BIBREF23": { | |
| "ref_id": "b23", | |
| "title": "How transferable are features in deep neural networks?", | |
| "authors": [ | |
| { | |
| "first": "Jason", | |
| "middle": [], | |
| "last": "Yosinski", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Jeff", | |
| "middle": [], | |
| "last": "Clune", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Yoshua", | |
| "middle": [], | |
| "last": "Bengio", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Hod", | |
| "middle": [], | |
| "last": "Lipson", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2014, | |
| "venue": "Advances in neural information processing systems", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "3320--3328", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. 2014. How transferable are features in deep neural networks? In Advances in neural information processing systems, pages 3320-3328.", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "FIGREF0": { | |
| "num": null, | |
| "type_str": "figure", | |
| "uris": null, | |
| "text": "High-level system architecture. Acoustic emotion recognition system uses neural speech representations for affective state classification. the speaker is not directly observable." | |
| }, | |
| "FIGREF1": { | |
| "num": null, | |
| "type_str": "figure", | |
| "uris": null, | |
| "text": "Architecture of the ASR model used in this work, following the DeepSpeech 2 architecture." | |
| }, | |
| "FIGREF2": { | |
| "num": null, | |
| "type_str": "figure", | |
| "uris": null, | |
| "text": "(a) Baseline speech emotion recognition (SER) model, containing two bi-directional GRU layers (b) A variant of the fine-tuning network (Fine-tuning MP-1) which uses the temporal pooled representations of the first recurrent layer of the ASR network. (c) A variant of the fine-tuning network (Fine-tuning RNN-1) which uses hidden memory representations of the first recurrent layer of the ASR network as input to a new emotion-specific GRU layer. (d) The progressive network combines representations from the emotion recognition path with the temporal pooled first layer representations of the ASR network (here Progressive net 1). A concatenated vector is fed into the softmax layer for the final emotion classification. The ASR branch of the network remains static and is not tuned during backpropagation in (b), (c) and (d)." | |
| }, | |
| "FIGREF3": { | |
| "num": null, | |
| "type_str": "figure", | |
| "uris": null, | |
| "text": "Confusion matrices for our best baseline, Fine-tuning MP and Progressive models averaged over 10-folds on the IEMOCAP dataset." | |
| }, | |
| "FIGREF4": { | |
| "num": null, | |
| "type_str": "figure", | |
| "uris": null, | |
| "text": "Most frequent class 0.39 \u00b1 0.04 0.33 \u00b1 0.2 NaN Random prediction 0.25 \u00b1 0.02 0.25 \u00b1 0.02 0.24 \u00b1 0.02 Baseline 64 units 0.52 \u00b1 0.04 0.54 \u00b1 0.06 0.48 \u00b1 0.05 Baseline 96 units 0.53 \u00b1 0.01 0.55 \u00b1 0.03 0.51 \u00b1 0.02 Baseline 128 units 0.50 \u00b1 0.04 0.51 \u00b1 0.06 0.47 \u00b1 0.05 Fine-tuning MP-1 0.46 \u00b1 0.05 0.52 \u00b1 0.07 0.31 \u00b1 0.11 Fine-tuning MP-2 0.54 \u00b1 0.04 0.54 \u00b1 0.06 0.52 \u00b1 0.04 Fine-tuning MP-3 0.55 \u00b1 0.02 0.56 \u00b1 0.03 0.53 \u00b1 0.03 Fine-tuning RNN-1 0.53 \u00b1 0.04 0.57 \u00b1 0.04 0.48 \u00b1 0.08 Fine-tuning RNN-2 0.57 \u00b1 0.03 0.59 \u00b1 0.05 0.56 \u00b1 0.03 Fine-tuning RNN-3 0.56 \u00b1 0.02 0.57 \u00b1 0.04 0.55 \u00b1 0.03 Progressive net-1 0.56 \u00b1 0.02 0.57 \u00b1 0.04 0.55 \u00b1 0.03 Progressive net-2 0.58 \u00b1 0.03 0.61 \u00b1 0.04 0.57 \u00b1 0.03 Progressive net-3 0.57 \u00b1 0.03 0.59 \u00b1 0.03 0.56 \u00b1 0.04" | |
| } | |
| } | |
| } | |
| } |