Speech and Computer: 23rd International Conference, SPECOM 2021, St. Petersburg, Russia, September 27–30, 2021, Proceedings (Lecture Notes in Computer Science, 12997)

This book constitutes the proceedings of the 23rd International Conference on Speech and Computer, SPECOM 2021, held in St. Petersburg, Russia, in September 2021.* The 74 papers presented were carefully reviewed and selected from 163 submissions. The papers present current research in the area of computer speech processing including audio signal processing, automatic speech recognition, speaker recognition, computational paralinguistics, speech synthesis, sign language and multimodal processing, and speech and language resources. *Due to the COVID-19 pandemic, SPECOM 2021 was held as a hybrid event. SPECOM 2021 Preface Organization Contents Text-Independent Speaker Verification Employing CNN-LSTM-TDNN Hybrid Networks 1 Introduction 1.1 Related Work 1.2 Our Contributions 2 Proposed Architecture 2.1 SpecAugment and CNN-based Feature Extraction 2.2 TDNN-LSTM-based Frame-Level Network 2.3 Multi-level Statistics Pooling 3 Experiments 3.1 Experimental Setups 3.2 Experiments on NIST SRE 2016 3.3 Experiments on Voxceleb 3.4 Experiments on Task 2 of SdSV Challenge 2021 4 Conclusion References End-to-End Voice Spoofing Detection Employing Time Delay Neural Networks and Higher Order Statistics 1 Introduction 1.1 Related Work 1.2 Our Contributions 2 Spoofing Detection Using TDNN and HOS 2.1 Higher Order Statistics (HOS) 2.2 Additive Angular Margin (AAM) Softmax 3 Experiments 3.1 ASVspoof2019 Challenge Corpora 3.2 Evaluation Metrics 3.3 Experimental Setup 3.4 Results 4 Conclusion References Assessing Velar Gestures Timing in European Portuguese Nasal Vowels with RT-MRI Data 1 Introduction 2 Methods 2.1 Articulatory Data 2.2 Post-processing and Determination of Articulatory Events 2.3 Timings 3 Results 4 Discussion 5 Conclusion References Designing and Deploying an Interaction Modality for Articulatory-Based Audiovisual Speech Synthesis 1 Introduction 2 Background and Related Work 2.1 AudioVisual Speech Synthesis 2.2 Articulatory-Based Audiovisual Synthesis 2.3 Multimodal Interaction 3 Evolutions of the Articulatory-Based Audiovisual Synthesis Core 4 Interaction Modality for Audiovisual Speech Synthesis 4.1 Audiovisual Speech Synthesis Output Modality 4.2 Sample Text Input Modality 4.3 Test Application 5 Conclusion References Kurdish Spoken Dialect Recognition Using X-Vector Speaker Embedding 1 Introduction 2 Data Collection 3 Proposed Method 4 Results and Discussion 5 Conclusion References An ASR-Based Tutor for Learning to Read: How to Optimize Feedback to First Graders 1 Introduction 2 Research Background 3 Present Study 3.1 Aim of the Study 3.2 The Reading Tutor 4 Method 4.1 Transcribed Speech Data 4.2 Kaldi ASR Systems Pipeline 4.3 ASR Systems 4.4 Procedure for Performance Evaluation of ASR Systems 5 Results 5.1 Performance Evaluation of Different Systems 5.2 Performance on Different Reading Tasks 5.3 Performance on Function and Content Words 6 Discussion and Conclusion References Velocity Differences Between Velum Raising and Lowering Movements 1 Introduction 2 Method 2.1 Subjects and Corpus 2.2 Experimental Setup and Recording Procedure 2.3 Tracking of the Lateral Pharyngeal Wall Movement 2.4 Determination of Movement Velocities 3 Results 4 Discussion and Conclusion References Pragmatic Markers of Russian Everyday Speech: Invariants in Dialogue and Monologue 1 Introduction 2 Typology of Pragmatic Markers 3 The Dialogues Sample (ORD Corpus) 4 The Monologues Sample (SAT Corpus) 5 Research Data 6 Statistics of Basic PM (Invariants) 7 Conclusion References Language Adaptation for Speaker Recognition Systems Using Contrastive Learning 1 Introduction 2 ResNet-Based Speaker Embedding 3 Domain Adaptation Approaches 3.1 Multilingual Learning 3.2 Transfer Learning 3.3 Contrastive Learning 4 Experiments and Results 4.1 Experimental Protocol 4.2 Results 5 Conclusion References Evaluating X-Vector-Based Speaker Anonymization Under White-Box Assessment 1 Introduction 2 Anonymization Technique 2.1 The Voice Conversion System 2.2 Design Choices for Anonymization 3 Experimental Setup 3.1 Dataset 3.2 Utility and Privacy Metrics 3.3 Evaluation Methodology 4 Experimental Results 4.1 Global Results 4.2 Detailed Analysis 4.3 Utility Results 5 Conclusion References Improved Prosodic Clustering for Multispeaker and Speaker-Independent Phoneme-Level Prosody Control*-10pt 1 Introduction 1.1 Related Work 1.2 Proposed Method 2 Method 2.1 Multispeaker Model 2.2 Dataset Augmentation 2.3 F0 Normalization 2.4 Balanced Duration Clustering 2.5 Speaker Adaptation 3 Experiments and Results 3.1 Objective Evaluation 3.2 Subjective Evaluation 4 Conclusion References Initial Experiments on Question Answering from the Intrinsic Structure of Oral History Archives 1 Introduction 2 Related Work 3 Approach 3.1 Creating the Mined MALACH QA Corpus 3.2 Transformer Neural Networks 3.3 Experimental Setup 4 Results and Discussion 5 Conclusion and Future Work References Imagined, Intended, and Spoken Speech Envelope Synthesis from Neuromagnetic Signals 1 Introduction 2 Data Collection 3 Methods 3.1 Data Preparation 3.2 Regression Model 3.3 Comparison of Brainwaves for Regression 4 Results and Discussion 4.1 Regression Performance 4.2 Comparison with Low Frequency Brainwaves 5 Conclusion References What Causes Phonetic Reduction in Russian Speech: New Evidence from Machine Learning Algorithms 1 Introduction 2 The Previous Study: What Went Wrong? 3 Random Forest, Gradient Boosting, Perceptron and Feature Selection 3.1 Working with Single Words 3.2 Working with Multiword Units 4 Discussion and Conclusion References Toxic Comment Classification Service in Social Network 1 Introduction 2 Related Work 2.1 Content Moderation of User Generated Content 2.2 Text Classification 3 Data and Preprocessing 4 Features 4.1 Vector Models 4.2 Content Features 5 Experiments and Results 5.1 Classification of Toxic Comments 5.2 Content Features Selection 6 System Architecture 7 Conclusion References Deep Learning Based Engagement Recognition in Highly Imbalanced Data 1 Introduction 2 Related Work 2.1 Engagement Recognition in HCI and HRI 2.2 Engagement Recognition Using DAiSEE 3 Dataset Analysis 4 Methodology 4.1 EmoVGGFace2 4.2 VGGFace2-SA 4.3 Recurrent Block 4.4 Methods for Imbalance Elimination 5 Experimental Results 6 Discussion 7 Conclusion References Intraspeaker Variability of a Professional Lecturer: Ageing, Genre, Pragmatics vs. Voice Acting (Case Study) 1 Introduction 2 Methodology 2.1 Data 2.2 Measurements 3 Results 3.1 The Ageing Impact on Prosodic Variability 3.2 The Genre Impact on Prosodic Variability 3.3 The Pragmatics Impact on Prosodic Variability 3.4 The Prosodic Variability of Voice Acting 4 Conclusion and Discussion References An Ensemble Approach for the Diagnosis of COVID-19 from Speech and Cough Sounds 1 Introduction 2 Related Work 3 Description of Our Adopted Approaches 3.1 Data Augmentation 3.2 Deep Supervised Feature Representations 3.3 End-to-End Diagnosis with ResNet 4 Experiments 4.1 Database 4.2 Baseline COVID-19 Diagnosis Systems 4.3 Evaluation Measure 4.4 Classifiers 4.5 Results and Discussion 5 Conclusion References Where Are We in Semantic Concept Extraction for Spoken Language Understanding? 1 Introduction 2 MEDIA Dataset 3 Overview of Approaches Proposed for the MEDIA Benchmark 3.1 Cascade Approach 3.2 End-to-End Approach 3.3 System Performance 4 Improving the State of the Art 4.1 BERT and CamemBERT Models 4.2 Wav2vec Models 4.3 Cascade Approach with Pre-trained Models 4.4 Results and Discussion 5 Conclusion References Learning Mizo Tones from F0 Contours Using 1D-CNN 1 Introduction 2 Previous Works and Motivation 3 Mizo Speech Corpus 4 Methodology 4.1 Estimation of F0 and Normalization 4.2 Proposed 1D-CNN-based Tone Detection System 4.3 Baseline Tone Detection System Using DNN 5 Results and Discussion 6 Conclusion References OCR Improvements for Images of Multi-page Historical Documents 1 Introduction 2 Related Work 3 Processing Pipeline 4 Page Segmentation 5 Deskew Algorithm 6 Optical Character Recognition 6.1 Synthetic Document Generation 6.2 Tesseract Retraining 7 Results and Ablation Study 8 Conclusion and Future Work References X-Bridge: Image-to-Image Translation with Reconstruction Capabilities 1 Introduction 2 Related Work 2.1 Pix2pix 2.2 UNIT/MUNIT 3 X-Bridge Method 4 Experiments 4.1 Dataset 4.2 Qualitative Results 4.3 Quantitative Results 5 Conclusion References Who is Selling to Whom – Feature Evaluation for Multi-block Classification in Invoice Information Extraction 1 Introduction 2 Related Work 3 The Method 3.1 Pre-processing and Annotations 3.2 Feature Extraction 3.3 The Classifier 4 Experiments and Evaluation 5 Conclusion and Future Directions References Multimodal Corpus Analysis of Autoblog 2020: Lecture Videos in Machine Learning 1 Introduction 2 Related Work 3 Corpus Description 4 Corpus Analysis 4.1 ASR Performance 4.2 Sentence Segmentation 4.3 Keyword Extraction 5 Discussion and Conclusion References Text and Synthetic Data for Domain Adaptation in End-to-End Speech Recognition 1 Introduction 2 Related Work 3 Method and Training 3.1 Baseline 3.2 Training and Domain Data 3.3 Synthesized Speech 3.4 Text-Encoder 4 Results 4.1 Basic Adaptation Methods 4.2 Comparison Synthetic Speech and Text-Encoder 4.3 Results with Text-Encoder 5 Conclusion References Speaker-Invariant Speech-to-Intent Classification for Low-Resource Languages 1 Introduction 2 Related Work 3 Methodology 4 Experiment 4.1 ASR Intermediate Representations 4.2 Self-supervised Features 4.3 Dataset 5 Results 6 Conclusion and Future Work References Speaker-Dependent Visual Command Recognition in Vehicle Cabin: Methodology and Evaluation 1 Introduction 2 Related Work 3 Data 4 Proposed Methodology 4.1 Visual Data Preprocessing and Labelling 4.2 Mouth Region Detection and Extraction 4.3 Model Training 5 Evaluation Experiments 6 Conclusion and Future Work References Optimised Code-Switched Language Model Data Augmentation in Four Under-Resourced South African Languages 1 Introduction 2 Dataset 3 Experimental Strategy 3.1 Metrics 3.2 Text Synthesis and N-Gram Augmentation 3.3 Hyperparameter Tuning 4 Experimental Results 4.1 Hyperparameter Tuning 4.2 Speech Recognition 5 Conclusion References Synthesis Speech Based Data Augmentation for Low Resource Children ASR 1 Introduction 2 Data 3 Baseline System 4 Data Augmentation 4.1 Prosody Based Data Augmentation 4.2 Speech Synthesis Based Data Augmentation 4.3 Language Model Sampling 5 Conclusion References End-to-End Russian Speech Recognition Models with Multi-head Attention 1 Introduction 2 Architecture of End-to-End Speech Recognition Model 3 Encoder-Decoder with Multi-head Attention 4 Results of Experiments on Russian Speech Recognition 5 Conclusion and Future Work References Word-Level Style Control for Expressive, Non-attentive Speech Synthesis 1 Introduction 1.1 Related Work 1.2 Our Contributions 2 Method 2.1 Word Style Encoder 2.2 Word Sequence Encoder 2.3 Prior Autoregressive Encoder 3 Experiments 3.1 Experimental Setup 3.2 Reconstruction Performance 3.3 Subjective Evaluation 3.4 Objective Assessment of Model's Generative Behavior 3.5 Interpretability of Tokens 3.6 Style Transfer 4 Conclusion References Perceiving Speech Aggression with and without Textual Context on Twitter Social Network Site 1 Introduction 1.1 Literature Review 1.2 Research Question and the Hypothesis 2 Methodology 2.1 Recruitment of Participants 2.2 Experiment Procedure 3 Results 3.1 Results of the First Research Step 3.2 Results of the Second Research Step 4 Conclusion References Assessing Speaker Interpolation in Neural Text-to-Speech 1 Introduction 2 Related Work 3 Prosody/Style Modelling 3.1 Vectors 3.2 Dynamic Layers 3.3 Several Places of Conditioning 4 Evaluation Methodology 4.1 Objective Evaluation 4.2 Subjective Evaluation 5 Setup 5.1 Architecture 5.2 Data 6 Results 6.1 Speakers 6.2 Objective 6.3 Subjective 7 Discussion 7.1 On Architecture and Hyperparameter Exploration 7.2 On Several Places of Conditioning 7.3 On Speaker Identity and Other Features in Text-to-Speech 8 Conclusion References A Mobile Application for Detection of Amyotrophic Lateral Sclerosis via Voice Analysis 1 The Relevance of Solving the Problem of Developing Affordable Non-invasive Methods for Early Diagnosis of ALS 2 Specificity of ALS Diagnostics by Voice Analysis 2.1 Speech Tasks Used in ASL Detection Systems 2.2 Acoustic Voice and Speech Features Used in ALS Detection Systems 2.3 Classification Methods Used in Systems for Detecting Speech Disorders in ALS 3 The Proposed Method for Evaluating the State of the Voice Function in Patients with ALS 3.1 The Used Acoustic Features 3.2 Jitter and Shimmer Parameters Optimization 3.3 Classifier Models 3.4 The Database of Voices 3.5 Train and Validation of Classifier Models 3.6 ALS Expert Mobile Application 4 Conclusion References Child’s Emotional Speech Classification by Human Across Two Languages: Russian & Tamil 1 Introduction 2 Methods 2.1 Speech Data Collection 2.2 Listeners 2.3 Data Analysis 3 Results 3.1 Characteristics of Stimulus Material 3.2 Perceptual Data: Spontaneous Speech vs Acting Speech 3.3 Acoustic Features of Emotional Speech 4 Discussion 5 Conclusion References Analysis of Dialogues of Typically Developing Children, Children with Down Syndrome and ASD Using Machine Learning Methods 1 Introduction 2 Datasets 3 Methods 4 Results & Discussion 5 Conclusion References Speaker Adaptation with Continuous Vocoder-Based DNN-TTS 1 Introduction 2 Related Work 3 Methodology 3.1 Speech Corpora 3.2 Continuous Vocoder 3.3 Build an Average Voice Model (AVM) 3.4 Adapt the AVM for the Adapt Speaker 4 Results 4.1 Objective Evaluation 4.2 Subjective Listening Test 5 Conclusion References Automatic Recognition of the Psychoneurological State of Children: Autism Spectrum Disorders, Down Syndrome, Typical Development 1 Introduction 2 Dataset 3 Description of the Schema for Automated Recognition 4 Experimental Results 5 Conclusion References Study on Acoustic Model Personalization in a Context of Collaborative Learning Constrained by Privacy Preservation 1 Introduction 2 Related Work 3 Model Adaptation 4 Experimental Setup 4.1 ASR System 4.2 Experimental Methodology 4.3 Datasets 5 Experimental Results 6 Conclusion References USC: An Open-Source Uzbek Speech Corpus and Initial Speech Recognition Experiments 1 Introduction 2 Related Work 3 The USC Dataset Construction 3.1 Crowdsourcing 3.2 Audiobooks 3.3 Dataset Statistics and Structure 4 Speech Recognition Experiments 4.1 Experimental Setup 4.2 Experiment Results 5 Conclusion References A Study of Multilingual End-to-End Speech Recognition for Kazakh, Russian, and English 1 Introduction 2 Related Work 3 Speech Recognition Models 3.1 Monolingual Model 3.2 Multilingual Model 4 Datasets 4.1 The Kazakh Language 4.2 The Russian Language 4.3 The English Language 5 Speech Recognition Experiments 5.1 Experimental Setup 5.2 Experiment Results 6 Discussion and Future Work 7 Conclusion References Dialog Speech Sentiment Classification for Imbalanced Datasets 1 Introduction 2 Related Work 3 Data 3.1 SWITCHBOARD-Sentiment Dataset 3.2 IEMOCAP-3 Dataset 4 Experiments and Results 4.1 Acoustic Sentiment Analysis 4.2 Text-Based Sentiment Analysis 4.3 Combination 4.4 Results 5 Conclusion References Explicit Control of the Level of Expressiveness in DNN-Based Speech Synthesis by Embedding Interpolation 1 Introduction 2 TTS Model from Prior Work 3 Proposed Idea 4 Data and Implementation Details 4.1 Database 4.2 Implemented TTS Model 5 Results 5.1 Subjective Evaluation 5.2 Objective Evaluation 6 Conclusions References Experimental Analysis of Expert and Quantitative Estimates of Syllable Recordings in the Process of Speech Rehabilitation 1 Introduction 2 Experiment 2.1 Data Description 2.2 Evaluation Algorithms 3 Results 4 Conclusion References Methods for Using Class Based N-gram Language Models in the Kaldi Toolkit 1 Introduction 1.1 Word Classes 1.2 The Kaldi Speech Recognition Toolkit 2 Relevant Work 3 Proposed Methods for Modifying the Kaldi Transducer 3.1 Implicit Expansion of Class N-Grams 3.2 Explicit Expansion of Class N-Grams 4 Testing and Conclusion References Spectral Root Features for Replay Spoof Detection in Voice Assistants 1 Introduction 2 Spectral Root Cepstrum 2.1 Speech Signal Modeling 2.2 Cepstrum Analysis: Logarithmic vs Spectral Root 2.3 Proposed SRCC Feature Set 3 Experimental Setup 3.1 Dataset 3.2 Feature Set, Classifier and Evaluation Metric 4 Experimental Results 4.1 Choice of 4.2 Spectrographic Analysis 4.3 Results 5 Summary and Conclusion References Influence of the Aggressive Internet Environment on Cognitive Personality Disorders (in Relation to the Russian Young Generation of Users) 1 Introduction 2 Method 3 Results 3.1 Clinical Case Studies. Conclusions on the Analysis of the Interview Content 3.2 Algorithm for Conclusion of No-Suicide Contract in Treatment of Patients with Internet Addiction Disorders 4 Conclusion References Media Content vs Nature Stimuli Influence on Human Brain Activity 1 Introduction 2 Method 3 Results 4 Conclusion References Can Your Eyes Tell Us Why You Hesitate? Comparing Reading Aloud in Russian as L1 and Japanese as L2 1 Introduction 1.1 Silent Pauses in Unprepared Reading as a Hesitation Phenomenon 1.2 Eye Tracking Studies of Reading in First and Second Languages 1.3 The Goals of the Study 2 Experiment 2.1 Material 2.2 Participants 2.3 Procedure 2.4 The Principles of Silent Hesitation Pauses Detection 2.5 The Principles of Eye Movements Analysis 2.6 Results 3 Discussion and Conclusions References Recognition of Heavily Accented and Emotional Speech of English and Czech Holocaust Survivors Using Various DNN Architectures 1 Introduction 2 Data Sources 2.1 English and Czech Parts of MALACH Archive 2.2 English and Czech LDC MALACH Data Used for Experimental Work 2.3 Annotation and Phonetic Transcription 2.4 Training and Test Subsets Preparation 3 Building LVCSR 3.1 Acoustic Feature Extraction 3.2 Acoustic Modeling 3.3 Language Modeling 3.4 Decoding 4 Results 5 Conclusions References Assessing Speaker-Independent Character Information for Acted Voices 1 Introduction 2 Character Representation Extraction and Evaluation 2.1 Character Similarity System Overview 2.2 Character-Oriented Representation 2.3 Voice Similarity Model 2.4 Corpus Description 2.5 Performance of the P-vector Representation 3 Estimation of the Amount of Character Information in the P-vector Representation 3.1 Random Association Protocol 3.2 Random Associations Subsets 3.3 Experiments and Results 4 Discussion 5 Conclusion and Future Work References Influence of Speaker Pre-training on Character Voice Representation 1 Introduction 2 Neural-Network-Based Character Voice Representation 2.1 Character-Oriented Representation 2.2 Decision 2.3 Sequence Extractor 3 Experimental Protocol 3.1 Corpora 3.2 Highlighting the Speaker-Independent Character Information 3.3 Checking the Generalization Abilities of the Character Representation 3.4 Pre-training of X-vector Model 3.5 Training of P-vector Neural Network 3.6 Evaluation 4 Reduce the Speaker Discriminative Information 5 Give More Power to the Character Classification 6 Conclusion References Opinion Classification via Word and Emoji Embedding Models with LSTM 1 Introduction 2 Opinion Classification of Uzbek Movie Reviews 2.1 Pre-processing 2.2 Feature Extraction 2.3 Opinion Classification 3 Experimental Setup 3.1 Dataset 3.2 Feature Extraction 3.3 Classification Algorithm 4 Experimental Results 5 Conclusion References An Equal Data Setting for Attention-Based Encoder-Decoder and HMM/DNN Models: A Case Study in Finnish ASR 1 Related Work 2 Data 3 Attention-Based System 4 HMM System 4.1 Correct Subword Handling in Lexicon FST 5 Results 6 Discussion 7 Conclusion References Speaker-Aware Training of Speech Emotion Classifier with Speaker Recognition 1 Introduction 2 Methodology 2.1 Dataset 2.2 Feature Extraction 2.3 Speaker-Aware Fine-Tuning 2.4 Complete Pipeline 3 Experiments 3.1 Speaker-Dependent vs Speaker-Independent Speech Emotion Recognition 3.2 Speaker-Aware Emotion Recognition 4 Conclusion References Neural Network Recognition of Russian Noun and Adjective Cases in the Google Books Ngram Corpus 1 Introduction 2 Data and Method 3 Result 4 Conclusion References Is It a Filler or a Pause? A Quantitative Analysis of Filled Pauses in Hebrew 1 Introduction 1.1 The Realization of FPs in Hebrew 1.2 Goal and Hypothesis 2 Data and Methods 2.1 Measurements 3 Results 3.1 Use of FPs, Words, and Silent Pauses 3.2 Correlations Between Attributes per Task 3.3 Correlations Between Attributes per Role 3.4 Correlations Between Speakers in the Same Task 4 Discussion References Modified Group Delay Function Using Different Spectral Smoothing Techniques for Voice Liveness Detection 1 Introduction 2 Features Used 2.1 Motivation for Modifying Group Delay Function 2.2 LP Spectrum-Based Smoothing 2.3 Spectral Root-Based Smoothing 3 Experimental Setup 3.1 Database Used 3.2 Baseline 3.3 Feature Set, Classifier, and Performance Metric 4 Experimental Results 5 Summary and Conclusions References Complex Rhythm Adjustments in Multilingual Code-Switching Across Mandarin, English and Russian 1 Introduction: An Overview of Rhythm Typology 2 Methodology 3 Results 3.1 Auditory Analysis of Pauses and Accents Delimiting IPs and Feet 3.2 Acoustic Analysis of Pitch and Tempo 4 Conclusion and Discussion References Increasing the Precision of Dysarthric Speech Intelligibility and Severity Level Estimate 1 Introduction 2 Methodology 2.1 Model and Experiments 3 Experimental Setup 3.1 Dataset 4 Results and Discussion 5 Conclusion References Articulation During Voice Disguise: A Pilot Study 1 Introduction 2 Phonology of Finnish and Russian 3 EMA Data Collection 3.1 Corpus Design 3.2 Recording Procedure 4 Methods 4.1 Functional Data Analysis 4.2 Automatic Speaker Verification 5 Results 5.1 Functional T-Tests 5.2 Speaker Verification Results 6 Conclusion References Improvement of Speaker Number Estimation by Applying an Overlapped Speech Detector 1 Introduction 2 Proposed Approach 2.1 Motivation 2.2 Speaker Counter 2.3 Overlapped Speech Detector 2.4 Fusion 3 Experimental Evaluation 3.1 Dataset Description 3.2 Signal Representations 3.3 Experimental Results 4 Conclusion References Mind Your Tweet: Abusive Tweet Detection 1 Introduction 2 Abusive Content 3 Related Work 4 Methodology 5 Dataset 6 Pre-processing 7 Experiments and Results 8 Conclusion and Future Work References Speaker Authorization for Air Traffic Control Security 1 Introduction 1.1 Voice Radio Communication in Air Traffic Control 1.2 Speaker Authorization 2 Proposed Approach 2.1 Architecture of the Speaker Authorization (SA) Module 2.2 Training Data 2.3 Test Data 2.4 Tests 2.5 Single Target Speaker Verification 2.6 Speaker Authorization/Speaker-Group Verification 2.7 Radio Channel Speaker Verification 3 Discussion and Conclusion References Prosodic Changes with Age: A Longitudinal Study on a Famous European Portuguese Native Speaker 1 Introduction 2 Related Work 3 Method 3.1 Speech Material and Data Annotation 3.2 Feature Extraction 3.3 Statistical Analysis 4 Result and Discussion 4.1 Intonation 4.2 Intensity 4.3 Rhythm 5 Conclusion References Automatic Selection of the Most Characterizing Features for Detecting COPD in Speech 1 Introduction 2 Methods 2.1 Acoustic Data 2.2 Data Pre-processing 2.3 Acoustic Features 2.4 Classification and Validation 2.5 Feature Selection 2.6 Evaluation of the Models 3 Results 4 Discussion 4.1 Limitations 4.2 Future Research 5 Conclusions References Multilingual Training Set Selection for ASR in Under-Resourced Malian Languages 1 Introduction 2 Languages 2.1 The Bambara Language 2.2 The Maasina Fulfulde Language 2.3 Scripts 3 Speech and Text Corpora 3.1 Monolingual Speech Corpora 3.2 Multilingual Speech Corpora 3.3 Text Corpora 4 Pronunciation Modelling 5 Language Modelling 6 Acoustic Modelling 7 Results and Discussion 8 Conclusions References Human and Transformer-Based Prosodic Phrasing in Two Speech Genres 1 Introduction 1.1 The Importance of the Topic 1.2 Current Objectives 2 Method 2.1 Data Description 2.2 Evaluation Measures 2.3 Experiment 1 2.4 Experiment 2 2.5 Experiment 3 3 Results 3.1 Experiment 1 – The General Model on LS 3.2 Experiment 2 – The General Model on NRS 3.3 Experiment 3 – Adaptation to a New Genre 3.4 Follow-Up Analyses 4 Discussion and Conclusion References Learning Efficient Representations for Keyword Spotting with Triplet Loss 1 Introduction 1.1 Previous Work on KWS 1.2 Previous Work on the Use of Triplet Loss for the Metric Embedding Learning 1.3 Our Contributions 2 Model Architectures 2.1 Input Preprocessing 2.2 Resnet Architecture 3 Experiments 3.1 Datasets and Tasks 3.2 Approach to Training Models 4 Results and Discussion References Regularized Forward-Backward Decoder for Attention Models 1 Introduction 2 Proposed Method 2.1 Attention Model 2.2 Adding a Backward Decoder 2.3 Regularization for Equal Sequence Lengths 2.4 Regularization for Unequal Sequence Lengths 3 Experiments 3.1 Training Details 3.2 Benchmark Details 3.3 Results 4 Conclusion References Induced Local Attention for Transformer Models in Speech Recognition 1 Introduction 2 Proposed Method 2.1 Transformer Network 2.2 Local Attention via Flexible Gaussian Window 2.3 Trainable Parameters of the Gaussian Mask 2.4 Global and Local Attention Score Fusion 3 Experiments 3.1 Training Setup 3.2 Ablation Study 3.3 Final Results 4 Conclusion References Applying EEND Diarization to Telephone Recordings from a Call Center 1 Introduction 2 Call Center Data Description 3 Speaker Diarization Based on X-Vectors 4 Diarization System Based on Identification 5 End-to-End Diarization System 5.1 EDA Calculation 6 Voice Activity Detector 7 Experiments 7.1 Training Data 7.2 Results 8 Discussion 9 Conclusion References Acoustic Characteristics of Speech Entrainment in Dialogues in Similar Phonetic Sequences 1 Introduction 2 Material 3 Method 4 Results 5 Conclusion References Predicting Biometric Error Behaviour from Speaker Embeddings and a Fast Score Normalization Scheme 1 Introduction 2 Speaker Embeddings 3 Biometric Menagerie 3.1 Goats 3.2 Wolves and Lambs 3.3 Sheep 4 Prediction of Biometric Error Behaviour 4.1 Annotation of Data 4.2 Regression Module 4.3 Score Normalization Module 5 Experimental Setup 5.1 X-Vector Embedding Training 5.2 Regressor Annotation and Training 5.3 Score Normalization 6 Results 6.1 Regression Results 6.2 Score Normalization Results 6.3 Normalization Speed 7 Conclusion and Future Work References Author Index This book constitutes the refereed proceedings of the 16th International Conference on Speech and Computer, SPECOM 2014, held in Novi Sad, Serbia. The 56 revised full papers presented together with 3 invited talks were carefully reviewed and selected from 100 initial submissions. It is a conference with long tradition that attracts researchers in the area of computer speech processing (recognition, synthesis, understanding etc.) and related domains (including signal processing, language and text processing, multi-modal speech processing or human-computer interaction for instance).