Prediction of Korean Prosodic Phrase Boundary by Efficient Feature Selection in Machine Learning

Hierarchical prosody structure generation is a key component for a speech synthesis system. One major feature of the prosody of Mandarin Chinese speech flow is prosodic phrase grouping. In this paper we proposed an approach for prediction of Chinese prosodic phrase boundaries in unrestricted Chinese text, which combines Conditional Random Fields (CRFs) model and TBL model. First a CRFs model is trained to predict the prosodic phrase boundaries. After that we apply a TBL based error driven learning approach to amend the initial prediction. A comparison is conducted between the new model and HMM for prosodic phrase break prediction. Experiments show that the combined approach improves overall performance. The precision and recall ratio are improved.

Hierarchical prosody structure generation is a key component for a speech synthesis system. One major feature of the prosody of Mandarin Chinese speech flow is prosodic phrase grouping. In this paper we proposed an approach for prediction of Chinese prosodic phrase boundaries from a limited amount of labeled training examples and some amount of unlabeled data using conditional random fields. Some useful unlabeled data are chosen based on the assigned labels and the prediction probabilities of the current learned model. The useful unlabeled data is then exploited to improve the learning. Experiments show that the approach improves overall performance. The precision and recall ratio are improved.

The paper analyzes both acoustic and linguistic features with different Chinese prosodic boundaries. Then a rule-learning approach was used to do the prosodic boundary labelling. In the paper the prosodic boundaries are classified into four levels, full intonational boundary with strong intonational marking with/without lengthening or change in speech tempo, prosodic phrase boundary with rather weak intonational marking, prosodic word boundary and phone foot boundary. Candidate acoustic and linguistic features related to prosodic boundary were extracted from the corpus to establish an example database. Based on this, a series of comparative experiments is conducted to collect the most effective features from the candidates. Results show that the selected candidates characterize the boundary features efficiently. Final experiments show that rule-learning approach introduced in the paper can achieve better prediction accuracy than the rule and RNN based methods and yet retain the advantage of the simplicity and understandability.KeywordsAcoustic ParameterLinguistic FeatureLinguistic InformationProsodic FeaturePhrase BoundaryThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.

Immediate integration of different types of prosodic information during on-line spoken language comprehension: An ERP study

To develop a hypothesis-free model that best predicts response to MTX drug in RA patients utilizing biologically meaningful genetic feature selection of potentially functional single nucleotide polymorphisms (pfSNPs) through robust machine learning (ML) feature selection methods. MTX-treated RA patients with known response were divided in a 4:1 ratio into training and test sets. From the patients' exomes, potential features for classifier prediction were identified from pfSNPs and non-genetic factors through ML using recursive feature elimination with cross-validation incorporating the random forest classifier. Feature selection was repeated on random subsets of the training cohort, and consensus features were assembled into the final feature set. This feature set was evaluated for predictive potential using six ML classifiers, first by cross-validation within the training set, and finally by analysing its performance with the unseen test set. The final feature set contains 56 pfSNPs and five non-genetic factors. The majority of these pfSNPs are located in pathways related to RA pathogenesis or MTX action and are predicted to modulate gene expression. When used for training in six ML classifiers, performance was good in both the training set (area under the curve: 0.855-0.916; sensitivity: 0.715-0.892; and specificity: 0.733-0.862) and the unseen test set (area under the curve: 0.751-0.826; sensitivity: 0.581-0.839; and specificity: 0.641-0.923). Sensitive and specific predictors of MTX response in RA patients were identified in this study through a novel strategy combining biologically meaningful and machine learning feature selection and training. These predictors may facilitate better treatment decision-making in RA management.

This paper describes the work done in improving the performance of Tamil speech recognition system by using Time Scale Modification (TSM) and Vocal Tract Length Normalization (VTLN) techniques. The speech recognition system for Tamil language was developed using a new approach of text independent speech segmentation, with a phoneme based language model for recognition. There is degradation in the performance of speech recognition due to variations in the speaking rate and vocal tract shape among different speakers. In order to improve the performance of speech recognition system, both TSM and VTLN normalization techniques were used in this work. The TSM was implemented using the Phase vocoder approach and the VTLN was implemented using speaker specific bark/mel scale in bark/mel domain. The performance of Tamil speech recognition system was improved by performing both TSM and VTLN normalization techniques.

Combined speech enhancement and auditory modelling for robust distributed speech recognition

Successful word learning in first language acquisition requires three components: segmentation of words from a continuous speech stream, identification of real-word referents or concepts, and mapping of linguistic label to meaning. This review presents behavioral evidence for how prosody is integral to each of these processes, with particular prosodic structures creating optimal word learning environments. Different prosodic strategies aid the three facets of word learning, with rhythm, stress, intonation, and phrasing interacting to facilitate learning. For example, recent behavioral work shows that word learning is enhanced when words appear at prosodic phrase boundaries and when new words carry a more prominent pitch accent. Finally, an integrative account of word learning is discussed in light of these findings.

This study advances a preliminary framework for conceptualising the prosodic nature and structure of Tiberian Hebrew (TH) represented by the ṭaʿămê hammiqrāʾ through an analysis of an extant Ashkenazi cantillation tradition. The ṭaʿămê hammiqrāʾ (lit. “the senses of the reading [viz. Scripture]”) are notations added by medieval scribes to the written text of the Hebrew Bible to preserve and transmit its oral performance. Modern prosodic theory and the musical concept of conjunct and disjunct melodic motion are used to demonstrate that the ṭaʿămê hammiqrāʾ have a highly structured iconic and intonational basis that organises the system and conforms substantially to Selkirk’s (2000, 2009) cross-linguistic prosodic hierarchy. The intonation-based prosodic model proposed in this study offers a solution to the limitation Dresher (1994, 2013; see also Dresher and DeCaen 2018; DeCaen and Dresher 2020) encounters with the intonational phrase domain of his prosodic model, permitting an alternative analysis of so-called pausal forms as lengthened forms, which can occur at prosodic phrase boundaries regardless of pause. The intonation-based model is tested by assessing how accurately it reflects the cross-linguistic prosodic features of restrictive and nonrestrictive relative clauses. The results indicate that TH distinguishes three prosodic classes of relatives—prosodically marked restrictives, prosodically marked nonrestrictives, and prosodically undifferentiated relatives—findings that accord with Birkner’s (2012) intonation-based study of the prosodic structure of German relative clauses.

Prosodic focus marking in Seoul Korean is known to be achieved primarily through prosodic phrasing, different from the use of prosody for this purpose in many other languages. This study investigates how children use prosodic phrasing for focus-marking purposes in Seoul Korean, compared to adults. Using a picture-matching game, we elicited semi-spontaneous production of SOV sentences in various focus conditions from monolingual Seoul Korean-speaking children aged 4 to 11 years. We found that the children varied prosodic boundaries to distinguish narrow focus from pre-focus and broad focus in a largely adult-like manner at the age of 4 to 5; at this age, they did not distinguish narrow focus from post-focus or contrastive focus using prosodic boundaries, similar to the adults. Their use of the prosodic boundaries in distinguishing the focus conditions was not fully adult-like in terms of frequency until the age of 10 to 11. In conjunction with the findings of previous studies on the acquisition of focus marking in Germanic languages, performed using a similar experimental method, our findings suggest that Seoul Korean-speaking children acquire the use of prosodic phrasing earlier than Dutch-speaking children acquiring the use of pitch accent but slightly later than Stockholm Swedish-speaking children acquiring the use of a prominence-marking high tone. These findings imply that the rate of focus-marking acquisition depends on the transparency of the form-meaning mapping between the phonological cue and focus.

Improved likelihood ratio test based voice activity detector applied to speech recognition

The emerging applications of wireless speech communication are demanding increasing levels of performance in noise adverse environments together with the design of high response rate speech processing systems. This is a serious obstacle to meet the demands of modern applications and therefore these systems often needs a noise reduction algorithm working in combination with a precise voice activity detector (VAD). This paper presents a new voice activity detector (VAD) for improving speech detection robustness in noisy environments and the performance of speech recognition systems. The algorithm defines an optimum likelihood ratio test (LRT) involving Multiple and correlated Observations (MO) and assuming a jointly Gaussian probability density function (jGpdf). An analysis of the methodology for N = {2, 3} shows the robustness of the proposed approach by means of a clear reduction of the classification error as the number of observations is increased. The algorithm is also compared to different VAD methods including the G.729, AMR and AFE standards, as well as recently reported algorithms showing a sustained advantage in speech/non-speech detection accuracy and speech recognition performance.

The performance of the speech recognition system for English classroom teaching is largely affected by the surrounding environment. These interference signals will seriously reduce the quality and intelligibility of the speech signal, thereby greatly reducing the performance of the far-field speech recognition system. Aiming at word order detection in English classroom teaching, this paper proposes an analysis model based on block coding and improved genetic algorithm. Moreover, for DNN-based single-channel speech enhancement algorithms, this paper proposes PDNNs and PLSTMs to solve the problem of serious performance degradation of prototype DNN speech enhancement under low signal-to-noise ratio. This method decomposes the entire enhancement task into multiple subtasks to complete, and the previously completed subtasks will provide prior knowledge for the subsequent subtasks, so that the subsequent subtasks can learn its goals better. In general, the experimental results prove the reliability of the model constructed in this paper.

Human Action Recognition (HAR) is a vital part of the healthcare sector. Medical practitioners are experiencing difficulty in recognizing human physiological conditions due to the enormous quantity of sensory stimulation. Machine learning techniques help to predict human physiological conditions, providing medical practitioners to work more efficiently. Feature selection is an essential part of discovering new knowledge in the majority of real-world problems when there are a lot of features. Feature selection is extremely useful because it speeds up decisions and enhances classification performance. The importance of feature selection in machine learning is dimension reduction in a massive multivariate data collection. This paper presents an effective feature selection method known as the Enhanced Recursive Feature Elimination (EFRE) for selecting key features from the data set for HAR prediction. The experimental results reveal that the ERFE technique selects the most suitable features for HAR prediction. The performance of the proposed ERFE approach is tested using different performance evaluation metrics. The performance analysis shows that the ERFE method outperforms existing feature selection methods with 88% accuracy.

Temporal information extraction is a popular and interesting research field in the area of Natural Language Processing (NLP). In this paper, we report our works on temporal relation identification within the TimeML framework. We worked on TempEval-2007 Task B that involves identification of relations between events and document creation time. Two different systems, one based on machine learning and the other based on handcrafted rules, are developed. The machine learning system is based on Conditional Random Field (CRF) that makes use of only some of the features available in TimeBank corpus in order to infer temporal relations. The second system is developed using a set of manually constructed handcrafted rules. Evaluation results show that the rule-based system performs better compared to the machine learning based system with the precision, recall and F-score values 75.9%, 75.9% and 75.9%, respectively under the strict evaluation scheme and 77.1%, 77.1% and 77.1%, respectively under the relaxed evaluation scheme. In contrast, CRF based system yields precision, recall and F-score values 74.1%, 73.6% and 73.8%, respectively under the strict evaluation scheme and 75.1%, 74.6% and 74.8%, respectively under the relaxed evaluation scheme.