---
title: "Dynamic VISC Features Improve Japanese Vowel Representation Across Whispered and Normal Phonation"
id: "plos-one-14-robust-acoustic-representation-of-japanese-vowels-under-phonation-variability"
canonical_url: "https://medichelpline.com/clinical-feed/plos-one-14-robust-acoustic-representation-of-japanese-vowels-under-phonation-variability"
content_type: "clinical_feed_article"
specialty: "Neurology"
source_name: "PLOS ONE (Medicine)"
source_url: "https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361125"
published_at: "2026-10-23T14:00:00.000Z"
evidence_level: "Journal Feed"
license: "CC-BY-NC-4.0 / Informational Use"
---
# Dynamic VISC Features Improve Japanese Vowel Representation Across Whispered and Normal Phonation
## Provenance & Clinical Metadata
- **Canonical URL:** https://medichelpline.com/clinical-feed/plos-one-14-robust-acoustic-representation-of-japanese-vowels-under-phonation-variability
- **Specialty:** [Neurology](https://medichelpline.com/clinical-feed/neurology.md)
- **Primary Source:** PLOS ONE (Medicine)
- **Source URL:** [Original Journal Publication](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361125)
- **Published At:** 2026-10-23T14:00:00.000Z
- **Evidence Rating:** Journal Feed
## Executive GIST (TL;DR)
- The study compared conventional **static formant** descriptors (F1, F2 midpoints) with dynamic spectral features derived from **VISC** (vector length and spectral angle) for Japanese vowels across normal and whispered phonation. - Investigators sampled 2410 manually verified vowel tokens drawn from 60 stimuli produced by 30 speakers in the JVS corpus; data included speaker ID, sex, and phonation condition labels. - The analysis treated Support Vector Machine (SVM) and Random Forest (RF) classifiers as analytical probes to evaluate representation stability, not as targets of classifier optimization. - Results indicated that incorporating low-dimensional dynamic spectral information produced more consistent vowel discrimination across phonation types than static-only measures, with similar trends observed for both SVM and RF. - The authors reported that classification differences appeared driven primarily by acoustic representation properties rather than model choice, supporting VISC’s relative robustness under phonation variability. - Data and all acoustic measurements (midpoint F1/F2 and VISC parameters) were deposited in figshare; the JVS corpus served as the publicly available source dataset. - The manuscript noted methodological constraints in prior work (single classifier, small homogeneous datasets) and addressed them by using two classifier frameworks and a larger, manually segmented token set; remaining limitations and unanswered questions were explicitly reported by the authors.
## Clinical Analysis & Structured Key Points
Robust acoustic representation of Japanese vowels under phonation variability using dynamic spectral analysis | PLOS One Browse Subject Areas ? Click through the PLOS taxonomy to find articles in your field. For more information about PLOS Subject Areas, click here . Article Authors Metrics Comments Media Coverage Reader Comments Figures Figures Abstract Phonation-induced acoustic variability, particularly in whispered speech, poses persistent challenges for stable vowel representation. Conventional static formant-based descriptors often fail to capture the dynamic spectral behavior of vowels under such conditions. This study evaluates the robustness of acoustic feature representations for Japanese vowels across normal and whispered phonation, comparing static formant measures with dynamic spectral features derived from Vowel Inherent Spectral Change (VISC). Dynamic features are used to characterize continuous formant trajectory patterns, while Support Vector Machine (SVM) and Random Forest (RF) classifiers serve as analytical probes to assess representation stability rather than as targets of algorithmic optimization. Results demonstrate that incorporating dynamic spectral information leads to more consistent vowel discrimination across phonation types compared to static-only representations, with similar performance trends observed across classification frameworks. These findings suggest that differences in classification outcomes are primarily driven by properties of the acoustic representations rather than by model choice. The study highlights the im-portance of representation-level design for robust speech signal analysis under phonation-induced variability and supports the use of low-dimensional dynamic features for vowel modeling in variable phonation conditions. Citation: Zhang Y, Cao T, Liu Y (2026) Robust acoustic representation of Japanese vowels under phonation variability using dynamic spectral analysis. PLoS One 21(10): e0361125. https://doi.org/10.1371/journal.pone.0361125 Editor: Laura Morett, University of Missouri Columbia, UNITED STATES OF AMERICA Received: February 24, 2026; Accepted: October 6, 2026; Published: October 23, 2026 Copyright: © 2026 Zhang et al. This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability: The data utilized in this study were sourced from the Japanese Versatile Speech (JVS) Corpus which is a freely accessible, publicly available dataset. Further details can be found on the introductory webpage at https://sites.google.com/site/shinnosuketakamichi/research-topics/jvs_corpus . All acoustic measurements generated in this study have been deposited in figshare. The dataset includes static midpoint F1 and F2 values, VISC parameters (vector length and spectral angle) for all vowel tokens, as well as corresponding speaker information (participant ID and sex) and phonation condition labels. The dataset is available at https://figshare.com/articles/dataset/Formant_F1_F2_and_VISC_Features_of_Japanese_Whispered_and_Normal_Vowels/33307449 . DOI: 10.6084/m9.figshare.33307449 . Funding: This work was supported by National Social Science Fund of China (Grant No. 25BYY150). The funder provided financial support only and had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. No additional external funding was received for this study. Competing interests: The authors have declared that no competing interests exist. 1. Introduction Speech processing research fundamentally relies on acoustic representations to capture linguistically relevant information from complex speech signals [ 1 – 3 ]. Across a wide range of analytical contexts, the effectiveness of speech analysis depends not only on modeling techniques, but critically on how speech signals are encoded and structured in acoustic space [ 4 , 5 ]. While substantial progress has been made in modeling normally phonated speech, acoustic representations are often implicitly assumed to remain stable across different speech production conditions. This assumption, however, has not been sufficiently examined, despite long-standing evidence that changes in speech production mechanisms can substantially alter spectral structure and formant behavior [ 6 , 7 ]. Variability arising from non-normal phonation therefore raises fundamental questions about whether commonly used acoustic representations preserve their descriptive adequacy when phonation conditions deviate from normal speech [ 8 ]. Understanding the robustness of acoustic representations under production-induced variability remains a foundational issue for speech signal analysis. Whispered and normal phonation differ substantially in both perceptual and acoustic properties. Whispered speech is produced through turbulent airflow without periodic vocal fold vibration, resulting in the absence of a fundamental frequency and a reduced harmonic structure [ 9 ]. These characteristics complicate acoustic analysis and pose persistent challenges for speech processing systems primarily designed for normal phonation. However, existing studies rarely address whether acoustic representations themselves remain stable under phonation-induced variability, leaving the mechanisms underlying representation robustness largely unexplored. Most prior work has evaluated whispered speech primarily in terms of classification performance, without explicitly examining whether commonly used acoustic representations remain stable when phonation conditions change [ 10 – 12 ]. As feature behavior may vary across phonation types and linguistic systems, commonly used acoustic descriptors may not preserve stable vowel representations under changing phonation conditions, and conclusions drawn solely from performance-based evaluations may therefore lack generalizability across languages or phonation modes. This issue is particularly salient for pitch-accent languages such as Japanese, which exhibit distinct prosodic and typological characteristics and have received comparatively limited attention in existing whispered speech research. To address these issues, the present study examines how phonation-related variability affects vowel representation in acoustic space, with particular attention to languages with distinct prosodic structures. Focusing on Japanese as a representative pitch-accent language, vowel representations under normal and whispered phonation are analyzed using both static formant-based descriptors and dynamic spectral features. Statistical classifiers, including Support Vector Machines (SVM) and Random Forests (RF), are employed as analytical probes to evaluate the robustness of different acoustic representations across phonation conditions, rather than as targets of recognition optimization. Within this framework, the study investigates how phonation-induced variability influences acoustic feature behavior and the stability of vowel representations, with the aim of elucidating representation-level properties relevant to speech analysis under non-normal phonation. Methodologically, this study treats statistical classifiers as analytical probes for examining acoustic representations, not as objects of algorithmic comparison. Differences in classifier performance are interpreted as reflections of feature behavior across phonation conditions, rather than only as indicators of recognition effectiveness per se. Within this framework, vowel representations under normal and whispered phonation are analyzed by comparing conventional static formant-based features with dynamic spectral features derived from Vowel Inherent Spectral Change (VISC). This design enables phonation-induced variability to be evaluated at the level of acoustic representation, without reliance on model-specific optimization. Specifically, this study addresses the following research questions: (1) How does phonation-induced variability reshape vowel representations in acoustic space when static formant measures are compared with dynamic spectral features? (2) To what extent are observed patterns of vowel representation stability consistent across different classification frameworks when dynamic features are incorporated? (3) How can classifier behavior be interpreted as evidence of representation robustness when evaluating low-dimensional acoustic features under small-sample conditions? Based on prior findings on dynamic spectral behavior, phonation-induced variability is expected to differentially affect static and dynamic acoustic representations, with VISC-based features exhibiting greater stability across phonation types. Although absolute classification outcomes may vary across modeling frameworks, stable performance trends associated with dynamic features are hypothesized to reflect properties of the acoustic representations themselves rather than model-specific effects. Under this view, dynamic spectral features are expected to provide more consistent encoding of vowel structure in the presence of phonation-induced variability. 2. Literature review The fundamental distinction between whispered and normal speech lies in the vibratory behavior of the vocal folds, which shifts from a closed, voiced state to a more open, voiceless configuration. In whispered speech, airflow passes through the vocal folds with reduced impedance, and frication replaces vocal fold vibration as the primary excitation source. This alteration in excitation mechanism substantially changes the acoustic characteristics of vowels, most notably through the attenuation or absence of fundamental frequency and a reduction in overall acoustic energy [ 12 , 13 ]. Additionally, because whispered speech requires greater articulatory precision to maintain intelligibility, it often exhibits a slower articulation rate and longer segmental duration than normally phonated speech [ 14 ]. The loss of salient acoustic cues such as fundamental frequency and harmonic structure poses significant challenges for speech analysis and modeling approaches developed primarily for normal phonation. Nevertheless, these phonation-induced differences establish an essential basis for examining how vowel representations vary—and potentially destabilize—across speech production conditions. Despite substantial progress in speech recognition, machine learning-based classification of whispered speech remains relatively underdeveloped, particularly with respect to generalizability across phonation types and languages. Mature statistical learning algorithms such as RF [ 15 ] and SVM [ 16 ] have been widely adopted in speech processing due to their distinct modeling characteristics. Existing studies suggest that classifier behavior is sensitive to both acoustic conditions and the structure of feature representations. RF has often been applied in small-sample settings and in feature-rich or noisy environments [ 17 ], whereas SVM is commonly used for high-dimensional representations, although its behavior may be more affected by noise and data imbalance [ 18 ]. More specifically, RF has been reported to handle multi-class classification with relatively limited preprocessing, while SVM tends to perform stably when acoustic features exhibit clearer class separability, with outcomes depending on kernel choice and regularization strategies [ 19 , 20 ]. Hybrid approaches that integrate RF-based feature selection with SVM classification have also been explored as a means of reducing redundancy in high-dimensional feature spaces [ 21 ]. Importantly, under non-normal phonation conditions, differences in classifier behavior are more plausibly interpreted as reflecting sensitivity to phonation-induced acoustic variability in feature representations, rather than intrinsic superiority of a given algorithm. The absence of key acoustic cues such as fundamental frequency in whispered speech introduces additional variability that challenges the stability of acoustic representations and, in turn, influences how classifiers respond to those representations. In this context, observed variation in algorithmic outcomes cannot be straightforwardly attributed to model effectiveness without considering the representational consequences of phonation change. Moreover, most existing whispered speech studies have focused on non-tonal languages (e.g., English) or tonal languages (e.g., Mandarin Chinese), whereas pitch-accent languages such as Japanese, which rely on distinct prosodic mechanisms involving funda-mental frequency, remain comparatively underexplored. This typological gap further motivates systematic evaluation of acoustic representations and their robustness across phonation types within pitch-accent language contexts. Existing research on whispered speech remains limited in the range of acoustic representations considered, a constraint that may affect the robustness of vowel modeling under variable phonation conditions. In studies of whispered Japanese speech, analysis has predominantly relied on static formant frequencies, particularly Formant 1 (F1) and Formant 2 (F2), measured at fixed temporal points within the vowel [ 22 ]. Empirical findings consistently show that whispered Japanese vowels exhibit elevated formant frequencies relative to normally phonated vowels, a pattern likewise reported for Chinese [ 23 ] and English [ 24 ]. While such static measurements effectively characterize broad spectral shifts associated with whispering, they provide only a partial account of vowel acoustics and do not fully reflect within-vowel spectral behavior. Evidence from perceptual and acoustic research demonstrates that temporal variation within vowels plays a central role in speech perception. Experimental manipulation of English vowels shows that removing the steady-state portion results in only a minimal reduction in intelligibility, from 93.1% to 92.4%, whereas eliminating dynamic onset and offset portions leads to a substantially larger decrease, to approximately 86.8% [ 25 ]. This asymmetry has been consistently replicated in subsequent studies [ 26 ]. In addition, vowels produced by non-native speakers often exhibit increased temporal variability, including greater dispersion in formant trajectories and prosodic cues, which can further influence intelligibility, particularly under whispered conditions [ 27 ]. Taken together, these findings indicate that static midpoint representations do not sufficiently characterize vowel behavior when phonation-related acoustic variability is present, underscoring the necessity of incorporating dynamic acoustic representations in the analysis of whispered speech. VISC is an acoustic representation designed to capture dynamic spectral variation within vowels and has been widely examined in studies of vowel perception and classification. In contrast to static features that describe formant characteristics at a single temporal point, typically the vowel midpoint, VISC characterizes spectral dynamics by analyzing formant trajectories at both the onset and offset of a vowel [ 25 ]. Dynamic spectral change is commonly characterized using two measures: vector length (VL), which captures the extent of formant displacement, and spectral angle (α), which reflects the overall direction of formant movement during the vowel [ 27 , 28 ]. Early studies demonstrated a strong correspondence between listeners’ perceptual identification of English vowels and confusion patterns derived from VISC-based formant measurements [ 29 ]. This relation-ship was subsequently supported by later investigations [ 30 ]. Even when substantial variation in surrounding consonantal context induces marked shifts in dynamic spectral patterns, vowel perception has been shown to remain predictable based on these dynamic features. These results point to the importance of formant information near vowel boundaries, which may be underrepresented when vowels are described solely by static mid-point values. A recent study examined the applicability of VISC to whispered vowel classification [ 31 ]. The authors conducted a preliminary investigation that refined the measurement procedure of VISC and evaluated its effectiveness relative to conventional static formant features across different phonation types and speaker genders. Using a SVM as the sole classification framework, the study reported that whispered vowels exhibited reduced vector length compared with normally phonated vowels, while spectral angle showed relatively limited variation across phonation types. In addition, the authors observed that dynamic features appeared less sensitive to speaker gender. When VISC features were combined with static formant measures, classification accuracy improved by approximately five percent compared with models relying on static features alone. Despite these initial findings, the extent to which the reported advantages of VISC can be interpreted as properties of the acoustic representation itself remains constrained by several methodological factors in the study [ 31 ]. Specifically, the use of a single classification algorithm restricted evaluation of the observed advantages of dynamic features across different modeling frameworks. Furthermore, the speech material consisted of repeated productions of a fixed text, yielding a relatively small and homogeneous dataset. Moreover, as VISC is derived solely from the trajectories of F1 and F2, it provides a comparatively low-dimensional acoustic description relative to feature sets such as Mel Frequency Cepstral Coefficients (MFCCs). Given these constraints, it remains unclear whether the observed benefits of VISC reflect intrinsic properties of the acoustic representation itself or are partially dependent on the specific classifier employed. These considerations motivate further examination of the stability of VISC-based vowel representations across different classification frameworks and phonation conditions, particularly under phonation-induced acoustic variability. 3. Materials and methods 3.1.Speech samples The data utilized in this study were sourced from the Japanese Versatile Speech (JVS) Corpus. This corpus is a freely accessible, publicly available dataset. Further details of which can be found on the introductory webpage mentioned in [ 32 ]. The corpus contains approximately 30 hours of professionally recorded speech from 100 native Japanese speakers, all of whom are trained voice actors. The corpus contains recordings in three phonation types: normal, whispered, and falsetto. Recordings were captured in a professional studio at a 48 kHz sampling rate, then downsampled to 24 kHz using the Speech Signal Processing Toolkit (SPTK), and saved in 16-bit RIFF WAV format. For this study, 60 stimuli were randomly selected from the five shared sentences, comprising recordings from 30 speakers (15 male and 15 female) across two phonation types (whispered and normal), ensuring coverage of all five Japanese vowel categories, as shown in Table 1 . This sampling strategy was designed to minimize the potential confounding effects of inter-speaker variability and to ensure consistent, high-quality manual vowel segmentation. The final dataset comprises 2410 manually verified vowel tokens across both phonation types, covering all five Japanese vowels in a range of consonantal environments. This sample provides a suitable basis for the supervised classification comparisons at the core of the study. Download: PNG larger image TIFF original image Table 1. Japanese speech stimuli and corresponding IPA transcriptions. Five sentences sourced from the JVS corpus, covering diverse phonetic contexts and including all five Japanese vowels. https://doi.org/10.1371/journal.pone.0361125.t001 As the ethics statement, this study conducted acoustic analysis based o
## Related Clinical Research

- [Deep multiscale CNN with temporal-channel attention for motor imagery EEG decoding](https://medichelpline.com/clinical-feed/plos-one-21-decoding-motor-imagery-eeg-signals-by-a-deep-learning-network-integrating.md)
- [Deep Learning 1D-CNN Accurately Classifies Nine Finger Movements from Surface EMG Signals](https://medichelpline.com/clinical-feed/plos-one-8-deep-learning-driven-fine-resolution-finger-movement-recognition-from-surface.md)
- [SPInfer: Fine-tuning Cellpose-SAM for scalable spatial proteomics in human brain tissue](https://medichelpline.com/clinical-feed/biorxiv-1-overcoming-segmentation-bottlenecks-in-brain-tissue-for-spatial-proteomic.md)
- [Cisplatin Modulates Honokiol Effects on Genes Governing Outer Hair Cell Function](https://medichelpline.com/clinical-feed/biorxiv-5-cisplatin-exposure-alters-the-effects-of-honokiol-on-genes-involved-in-outer.md)
- [Understanding the Complexity of Multiple Sclerosis Diagnosis: Lessons from Christina Applegate](https://medichelpline.com/clinical-feed/stat-news-0-opinion-christina-applegate-puts-a-face-on-one-of-the-most-consequential.md)

## Navigation
- [← Back to Neurology Feed](https://medichelpline.com/clinical-feed/neurology.md)
- [← All Clinical Specialties](https://medichelpline.com/clinical-feed.md)
## Medical & Regulatory Disclaimer

> [!CAUTION]
> MedicHelpline content is structured for research, educational, and professional discovery purposes. It does not constitute individual medical advice, clinical diagnosis, or treatment recommendations.
> Always verify dosing, contraindications, and regulatory alerts against official product labeling and primary regulatory sources before clinical decision-making.