Speech Corpus Research Articles

Comparison of Psychometric Functions Measured Using Remote Testing and Laboratory Testing.

The use of remote testing to collect behavioral data has been on the rise, especially after the COVID-19 pandemic. Here we present psychometric functions for a commonly used speech corpus obtained in remote testing and laboratory testing conditions on young normal hearing listeners in the presence of different types of maskers. Headphone use for the remote testing group was checked by supplementing procedures from prior literature using a Huggins pitch task. Results revealed no significant differences in the measured thresholds using the remote testing and laboratory testing conditions for all the three masker types. Also, the thresholds measured obtained in these two conditions were strongly correlated for a different group of young normal hearing listeners. Based on the results, excellent outcomes on auditory threshold measurements where the stimuli are presented both at levels lower than and above an individual's speech-recognition threshold can be obtained by remotely testing the listeners.

Creation of a diverse mixed-lingual emotional speech corpus with a framework for enhanced emotion detection

Speech Emotion Recognition (SER) has advanced considerably during the past 20 years. Till date, various SER systems have been developed for monolingual, multilingual and cross corpus contexts. However, in a country like India where numerous languages are spoken and often humans converse in more than one language, a dedicated SER system for mixed-lingual scenario is more crucial to be established which is the focus of this work. A self-recorded database that includes speech emotion samples with 11 diverse Indian languages has been developed. In parallel, a mixed-lingual database is formed with three popular standard databases of Berlin, Baum and SAVEE to represent mixed-lingual environment for western background. A detailed investigation of GeMAPS (Geneva Minimalistic Acoustic Parameter Set) feature set for mixed-lingual SER is performed. A distinct set of MFCC (Mel Frequency Cepstral Coefficients) coefficients derived from sine and cosine-based filter banks enriches the GeMAPS feature set and are proven to be robust for mixed-lingual emotion recognition. Various Machine Learning (ML) and Deep Learning (DL) algorithms have been applied for emotion recognition. The experimental results demonstrate GeMAPS features classified from ML has been quite robust for recognizing all the emotions across the mixed-lingual database of the western languages. However, with diverse recording conditions and languages of the Indian self-recorded database the GeMAPS with enriched features and classified using DL are proven to be significant for mixed-lingual emotion recognition.

Customized deep learning based Turkish automatic speech recognition system supported by language model.

In today's world, numerous applications integral to various facets of daily life include automatic speech recognition methods. Thus, the development of a successful automatic speech recognition system can significantly augment the convenience of people's daily routines. While many automatic speech recognition systems have been established for widely spoken languages like English, there has been insufficient progress in developing such systems for less common languages such as Turkish. Moreover, due to its agglutinative structure, designing a speech recognition system for Turkish presents greater challenges compared to other language groups. Therefore, our study focused on proposing deep learning models for automatic speech recognition in Turkish, complemented by the integration of a language model. In our study, deep learning models were formulated by incorporating convolutional neural networks, gated recurrent units, long short-term memories, and transformer layers. The Zemberek library was employed to craft the language model to improve system performance. Furthermore, the Bayesian optimization method was applied to fine-tune the hyper-parameters of the deep learning models. To evaluate the model's performance, standard metrics widely used in automatic speech recognition systems, specifically word error rate and character error rate scores, were employed. Upon reviewing the experimental results, it becomes evident that when optimal hyper-parameters are applied to models developed with various layers, the scores are as follows: Without the use of a language model, the Turkish Microphone Speech Corpus dataset yields scores of 22.2 -word error rate and 14.05-character error rate, while the Turkish Speech Corpus dataset results in scores of 11.5 -word error rate and 4.15 character error rate. Upon incorporating the language model, notable improvements were observed. Specifically, for the Turkish Microphone Speech Corpus dataset, the word error rate score decreased to 9.85, and the character error rate score lowered to 5.35. Similarly, the word error rate score improved to 8.4, and the character error rate score decreased to 2.7 for the Turkish Speech Corpus dataset. These results demonstrate that our model outperforms the studies found in the existing literature.

Automatic speaker and age identification of children from raw speech using sincNet over ERB scale

This paper presents the newly developed non-native children’s English speech (NNCES) corpus to reveal the findings of automatic speaker and age recognition from raw speech. Convolutional neural networks (CNN), which have the ability to learn low-level speech representations, can be fed directly with raw speech signals instead of using traditional hand-crafted features. Moreover, the filters that were learned using standard CNNs appeared to be noisy because they consider all elements of each filter. In contrast, sincNet can be able to generate more meaningful filters simply by replacing the first convolutional layer by a sinc-layer in standard CNNs. The low and high cutoff frequencies of the rectangular band-pass filter are the only parameters that can be learned in sincNet, which has the potential to extract significant speech cues from the speaker, such as pitch and formants. In this work, the sincNet model is significantly changed by switching from baseline Mel scale initializations to equivalent rectangular bandwidth (ERB) initializations, which has the added benefit of allocating additional filters in the lower region of the spectrum. Additionally, it needs to be highlighted that the novel sincNet model is well suited to identify the age of the children. The investigations on both read and spontaneous speech tasks in speaker identification, gender independent & dependent age-group identification of children outperform the baseline models with varying relative improvements in terms of accuracy.

Positional error distribution in Thai speakers’ acquisition of Korean stop consonants: A speech corpus analysis

Positional error distribution in Thai speakers’ acquisition of Korean stop consonants: A speech corpus analysis

Classification of Manners of the Prevocalic Alveolar Consonants in Machine Learning Using Dynamic Formant Transitions of Vowels in Korean Spontaneous Speech

The present study tried to classify the prevocalic manners of alveolar consonants using the formant transitions of the vowel in a Korean spontaneous speech corpus. Random forest and neural network models were trained and tested on selections of F1, F2, and F3 samples taken at the vowel onset and target of the vowel. It was found that prevocalic manners could not be manifested properly by the samples of any one or two formants, taken singly at the vowel onset or doubly at the vowel onset and target. Rather, both models trained on all the F1, F2, and F3 measurements taken doubly at the vowel onset and target, manifested the prevocalic manners robustly, though the random forest model outperformed the neural network model. The former model was further trained on additional predictors. Vowel identity facilitated model classification substan- tially more than F0, gender, speaking rate, and vowel duration. The contribution of the latter predictors was rather marginal.

Редукция на неударените гласни в съвременния български книжовен език: съпоставка на общоприети възгледи с корпусни данни

Bulgarian is often cited in phonological work for its vowel reduction, with the assumption that the six-vowel stressed inventory, /ɛ, a, ɔ, i, ɤ, u/, shrinks to a sub-inventory of three contrastive vowels in unstressed position, /i, ɤ, u/, by virtue of /ɛ, a, ɔ/ raising and merging with /i, ɤ, u/, respectively. The literature in Bulgarian, on the other hand, maintains that /ɛ–i/ do not merge in Standard Bulgarian; that unstressed vowels are more open in immediately pretonic syllables than elsewhere; that unstressed high vowels are lowered, while nonhigh vowels are raised; and that reduction is strongest in /a/. These claims have been challenged in recent work, and we present a new investigation based on 11 615 vowel tokens from 140 speakers in the BulPhonC speech corpus. MANOVA and LMM results provide clear evidence that there is no unstressed high-vowel lowering, no difference between pretonic vs other unstressed vowels, and that unstressed /a–ɤ/ and /ɔ–u/ merge completely through the raising of /a/ and /ɔ/ to [ɤ] and [u], respective-ly. At the same time, unstressed /ɛ–i/ remain spectrally distinct.

Disambiguation of Isolated Manipuri Tonal Contrast Word Pairs Using Acoustic Features

Manipuri is a low-resource, Tibeto-Burman tonal language spoken mainly in Manipur, a northeastern state of India. Tone identification is crucial to speech comprehension for tonal languages, where tone defines the word’s meaning. Automatic Speech Recognition for those languages can perform better by including tonal information from a powerful tone detection system. While significant research has been conducted on tonal languages like Mandarin, Thai, Cantonese, and Vietnamese, a notable gap exists in exploring Manipuri within this context. To address this gap, this work expands our previously developed handcrafted speech corpus, ManiTo, which comprises isolated Manipuri tonal contrast word pairs to study the tones of Manipuri. This extension includes contributions from 20 native speakers. Preliminary findings have confirmed that Manipuri has two unique tones, Falling and Level. The study then conducts a comprehensive acoustic feature analysis. Two sets of features based on Pitch contours, Jitter, and Shimmer measurements are investigated to distinguish the two tones of Manipuri. Support Vector Machine, Long Short-term Memory, Random Forest, and k-Nearest Neighbors are the classifiers adopted to validate the selected feature sets. The results indicate that the second set of features consistently outperformed the first set, demonstrating higher accuracy, particularly when utilizing the Random Forest classifier, which provides valuable insights for further advancements in speech recognition technology for low-resource tonal language Manipuri.

Classification patterns in conversational English fricatives: Between- and within- speaker analyses

This study examines between- and within- speaker patterns in random forest classification models for fricatives in conversational English. Prior investigations on the categorization of fricatives have mostly focused on group-level analyses and careful speech. We use a corpus of sociolinguistic interview speech from Western Canadian English, which represents a more casual speech style, and we compare group-level results to the findings of classification models on individual speakers’ productions. The models use 23 different spectral, temporal, and amplitudinal measures to predict phonetic labels of the fricatives. Our results reveal that certain top-ranked predictors at the group level (e.g., spectral peak frequency, segment duration) are also important for fricative classification in the individual models. Other measures (e.g., midpoint kurtosis, RMS) show substantial variability in their relative prominence and play greater roles in individual- versus group-level modeling. We discuss the implications of these findings for our understanding of individual variation in connected speech and what the findings may imply about models of production and perception.

Spectral properties of Quebec French sibilants

Most acoustic work on sibilants has focused on English /s/ and /ʃ/, examining spectra of the middle portion of these consonants. Voiced sibilants, which are rarer cross-linguistically (Ohala, 1983), as well as sibilants in other languages, have received less attention; in particular, there has been no previous work on Quebec French. Previous cross-linguistic studies have suggested important differences in sibilant acoustics exist between languages (Gordon et al., 2002), including in the dynamics of spectral properties over the course of the segment (Reidy, 2016). This study adds to the literature on the typology of sibilant acoustics by examining the spectral dynamics of sibilants in Quebec French (/s, ʃ, z, ʒ/). ∼28 k word-initial, pre-vocalic tokens from more than 100 speakers are extracted from a large corpus of parliamentary speech (Milne, 2014). For each token, multitaper spectra (Reidy, 2013) over a 20 ms window are calculated at 17 equidistant points. A variety of acoustic measures, including segment duration and spectral moments, are reported. Preliminary results examining static measures reveal that the anterior sibilants have higher spectral speak (∼5700 Hz vs ∼4600 Hz for posterior ones), as expected; unlike in English, however, / s/ and /ʃ/ have similar average duration.

Modern Standard Arabic speech disorders corpus for digital speech processing applications

Modern Standard Arabic speech disorders corpus for digital speech processing applications

Location of constriction in velar sounds in French

This study examines the constriction location (CL) of the velar stop [k] in French. While previous studies have investigated how vocalic contexts influence the CL of velar stops [Liker & Gibbon (2008) Clin. Ling. & Phon. 22(2); Tabain (2000) JPhon 28(2)], few investigated high-level changes in CL. Building on our prior work [Islam & Gick (2023) JASA 154], which showed an unexpected palato-velar articulation of [k] followed by [a] compared with [w] followed by [a], we tested whether this [k]-palatalizing is universal or specific to [a]. Using a French MRI speech corpus [Isaieva et al., Scientific Data 8], we measured constriction in [k] before [i], [ɛ], [o], [u], and [a]. MRI video frames were manually traced to mark the upper surface of the tongue and the lower surface of the hard palate, resulting in two contours. Using Euclidean distance, the location of the constriction was identified as the narrowest point and distance, respectively, between the two contours. A Python script that measured Euclidean distance from traced MRI frames calculated these distances from the MRI frames. Results indicate a frontal shift in [k] across all contexts, indicating a general fronting articulation of [k] in French speech.

The fundamental frequency (f0) distribution of American speakers in a spontaneous speech corpus*

The fundamental frequency (f0) distribution of American speakers in a spontaneous speech corpus*

Regional differences in the production of tones in standard mandarin

This study investigates the intricate patterns of tonal variations in monosyllabic words across distinct regional accents of Standard Mandarin, with a focus on Shanghai, Guangzhou, and Beijing. Due to the differing tonal systems of the local Chinese dialects in the three cities, speakers from these areas may exhibit distinct patterns in their production of tones in Standard Mandarin, influenced by their native Chinese dialects. Therefore, to capture the nuances of tonal differences, the study extracts the time-normalized f0 values of four tones (i.e., level, rising, dipping, and falling tones) from monosyllabic words in a Mandarin Chinese speech corpus. The overall f0 contours of the four tones within and across the three regional accents are modeled using growth curve analysis. We expect to find that regional accents of Standard Mandarin exhibit significant tonal variations such that the tone shapes of speakers from Shanghai and Guangzhou may significantly deviate from those of Beijing speakers. Specifically, there are variations in contour shape and temporal dynamics, particularly in the quadratic term for dipping tones and the slopes of rising and falling tones. These findings can provide valuable insights into the intricate characteristics of regional tonal variations in Standard Mandarin.

К проблеме интегрального описания речевых жанров

The article formulates the subject and tasks of the future speech-genre discipline – the integral description of speech genres – which, in the author’s opinion, claims to be central to the general theory of speech genres. The main sources of the theory of speech genres integrality include the high level of a speech genre as a speech and language unit, the complex, multidimensional and multicomponent nature of the theory of speech genres, the multidisciplinarity of the theory of speech genres (it draws on data from almost all the humanities and partly non-humanities). The integral description of speech genres will include the following parameters: the cultural and historical context of the speech genre; typological characteristics of speech genre (including their place in various typologies); speech-genre variantology and recurrence, including new technogenic and Internet derivatives from primary, traditional genres; representation in the corps. The integral description of speech genres and its parameters are discussed in connection with the division of linguistics (partly – other sciences, for example, literary criticism). It is shown that the majority of divisions of linguistics that are significant for the integral description of speech genres go back to Saussure’s idea of opposing internal and external linguistics, and the main tasks of the theory of speech genres correspond to those three “eternal” problems of linguistics formulated by V. M. Alpatov: “How is language structured?”, “How does language function?” and “How does language develop?” The author discusses directions of the most important genre studies, from which the integral description of speech genres should be “composed”, and the criteria used for their selection: 1) the model is already integral (multi-component and interdisciplinary); 2) universal (applicable to many languages and national cultures, in different historical periods); 3) based on this model, the largest number of studies was carried out. The first criterion is met by the model of T. V. Shmeleva, the second – by the universalist model of A. Wierzbicka, the third – by traditional speech-genre models: lexical, syntactic and pragmatic. In conclusion, the author discusses the unresolved problems of speech genres, which can be logically solved on the basis of the integral description of speech genres: typology (integral bases for the classification of speech genres claim to be more reliable), variantology, applied and experimental aspects (vocabulary representation of speech genres and corpus aspect, including the analysis of key phrases of speech genres).

MECOS: A bilingual Manipuri–English spontaneous code-switching speech corpus for automatic speech recognition

In this study, we introduce a new code-switched speech database with 57h of Manipuri–English annotated spontaneous speech. Manipuri is an official language of India and is primarily spoken in the north–eastern Indian state of Manipur. Most Manipur native speakers today are bilingual and frequently use code switching in everyday discussions. By carefully assessing the amount of code-switched speech in each video, recordings from YouTube are gathered. 21,339 utterances and 291,731 instances of code switching are present in the database. Given the code-switching nature of the data, a proper annotation procedure is used, and the data are manually annotated using the Meitei Mayek unicode font and the roman alphabets for Manipuri and English, respectively. The transcription includes the information of the speakers, non-speech information, and the corresponding annotation. The aim of this research is to construct an automatic speech recognition (ASR) system as well as offer a thorough analysis and details of the speech corpus. We believe that our research is the first to use an ASR system for Manipuri–English code-switched speech. To evaluate the performance, ASR systems based on hybrid deep neural network and hidden Markov model (DNN–HMM), time delay neural network (TDNN), hybrid time delay neural network and long short-term memory (TDNN–LSTM) and three end-to-end (E2E) models i.e. hybrid connectionist temporal classification and attention model (CTC-Attention), Conformer, wav2vec XLSR are developed for Manipuri–English language. In comparison to other models, pure TDNN produces outcomes that are clearly superior.

COMPARISON OF THE SPEECH TEXTS OF INDONESIAN PRESIDENT JOKO WIDODO AND PRESIDENT SUSILO BAMBANG YUDHOYONO: STUDY USING A CORPUS LINGUISTIC APPROACH

Purpose: This research has several objectives. First, determine lexical density and compare the lexical density. Second, to determine the key lexical density and compare the key lexical density. Third, to test the independence of the relationship between lexical variations and the text of President Joko Widodo's and President Susilo Bambang Yudhoyono's speeches. Theoretical Reference: The theoretical basis used in this research is the lexical analysis approach in linguistics. The application of lexical perspective analysis is expected to be able to review the communication used by each individual. The theoretical lexical discussion will also use a statistical independence analysis approach. The application of a statistical independence analysis approach is used to review a person's individual language abilities. Method: This research uses a qualitative and quantitative corpus linguistics approach. The corpus linguistic application used in this research is the KORTARA application (Korpus Nusantara). The research data is a corpus of 9 texts of President Joko Widodo's speeches and a corpus of 9 texts of President Susilo Bambang Yudhoyono which are official speeches every 16 August before the DPR of the Republic of Indonesia. Results and Conclusion: The results of this research reveal that the text corpus of President Joko Widodo's speech is richer and more varied than the text corpus of President Susilo Bambang Yudhoyono's speech in lexical use. This research also revealed that there is a relationship between lexical variation and the type of text of the President of the Republic of Indonesia's speech with a confidence level of 95%. The difference in lexical variation and frequency between the text corpus of President Joko Widodo's speech and the text corpus of Susilo Bambang Yudhoyono's speech is statistically significant at p < 0.05. Implication of Research: The implication of this research is the realization of the KORTARA corpus linguistic approach (Korpus Nusantara) which can facilitate research for small and large scale data. This research also reveals that the application of a statistical approach provides maximum results in the analysis of large-scale linguistic phenomena. Originality/value: The current study makes a valuable empirical contribution by combining statistical analysis using corpus and qualititative analysis to give comprehensive conclusion. This study is the answer toward the question about the reliability and validity of linguistic studies.

Denoising Convolutional Autoencoder Based Approach for Disordered Speech Recognition

Efficient assistive speech technology is essential for persons with cognitive disorders to improve their standard of life. Various kinds of cognitive disorders affect the speech articulation. Disordered Speech Recognition (DSR) can be used for rehabilitation and gain much importance as the disordered speakers population keeps increasing in recent years. The speech utterance is commonly represented in the form of spectrogram. Since the spectrograms are noisy and incomplete, the corresponding spectrograms need to be enhanced. We propose an approach that explores Denoising Convolutional Autoencoder (DCAE) to enhance the spectrograms of disordered speech utterances which are then utilized to train CNN to recognize the disordered words. Evaluation of proposed approach is carried out using the 20 words (acoustically similar word classes) dataset and 50 words dataset of Impaired speech corpus in Tamil and 100 common words dataset of UA-Speech database. Significantly better performance is achieved by proposed approach than HMM, DNN-HMM, LFMMI and CNN without enhancement. The spectrogram enhancement using DCAE helps to obtain better discrimination among overlapping disordered word classes and achieves a maximum Word Recognition Accuracy.

Evaluating the Efficacy of Traditional Machine Learning Models in Speaker Recognition: A Comparative Study Using the LibriSpeech Dataset

The efficacy of machine learning models in speaker recognition tasks is critical for advancements in security systems, biometric authentication, and personalized user interfaces. This study provides a comparative analysis of three prominent machine learning models: Naive Bayes, Logistic Regression, and Gradient Boosting, using the LibriSpeech test-clean dataset—a corpus of read English speech from audiobooks designed for training and evaluating speech recognition systems. Mel-Frequency Cepstral Coefficients (MFCCs) were extracted as features from the audio samples to represent the power spectrum of the speakers’ voices. The models were evaluated based on precision, recall, F1-score, and accuracy to determine their performance in correctly identifying speakers. Results indicate that Logistic Regression outperformed the other models, achieving nearly perfect scores across all metrics, suggesting its superior capability for linear classification in high-dimensional spaces. Naive Bayes also demonstrated high efficiency and robustness, despite the inherent assumption of feature independence, while Gradient Boosting showed slightly lower performance, potentially due to model complexity and overfitting. The study underscores the potential of simpler machine learning models to achieve high accuracy in speaker recognition tasks, particularly where computational resources are limited. However, limitations such as the controlled nature of the dataset and the focus on a single feature type were noted, with recommendations for future research to include more diverse environmental conditions and feature sets.

How Do Care Partners of People with Rare Dementia Use Language in Online Peer Support Groups? A Quantitative Text Analysis Study

We used quantitative text analysis to examine conversations in a series of online support groups attended by care partners of people living with rare dementias (PLWRD). We used transcripts of 14 sessions (>100,000 words) to explore patterns of communication in trained facilitators’ (n = 2) and participants’ (n = 11) speech and to investigate the impact of session agenda on language use. We investigated the features of their communication via Poisson regression and a clustering algorithm. We also compared their speech with a natural speech corpus. We found that differences to natural speech emerged, notably in emotional tone (d = −3.2, p < 0.001) and cognitive processes (d = 2.8, p < 0.001). We observed further differences between facilitators and participants and between sessions based on agenda. The clustering algorithm categorised participants’ contributions into three groups: sharing experience, self-reflection, and group processes. We discuss the findings in the context of Social Comparison Theory. We argue that dedicated online spaces have a positive impact on care partners in combatting isolation and stress via affiliation with peers. We then discuss the linguistic mechanisms by which social support was experienced in the group. The present paper has implications for any services seeking insight into how peer support is designed, delivered, and experienced by participants.