Formant Bandwidths Research Articles

ChildAugment: Data augmentation methods for zero-resource children's speaker verification.

The accuracy of modern automatic speaker verification (ASV) systems, when trained exclusively on adult data, drops substantially when applied to children's speech. The scarcity of children's speech corpora hinders fine-tuning ASV systems for children's speech. Hence, there is a timely need to explore more effective ways of reusing adults' speech data. One promising approach is to align vocal-tract parameters between adults and children through children-specific data augmentation, referred here to as ChildAugment. Specifically, we modify the formant frequencies and formant bandwidths of adult speech to emulate children's speech. The modified spectra are used to train emphasized channel attention, propagation, and aggregation in time-delay neural network recognizer for children. We compare ChildAugment against various state-of-the-art data augmentation techniques for children's ASV. We also extensively compare different scoring methods, including cosine scoring, probabilistic linear discriminant analysis (PLDA), and neural PLDA. We also propose a low-complexity weighted cosine score for extremely low-resource children ASV. Our findings on the CSLU kids corpus indicate that ChildAugment holds promise as a simple, acoustics-motivated approach, for improving state-of-the-art deep learning based ASV for children. We achieve up to 12.45% (boys) and 11.96% (girls) relative improvement over the baseline. For reproducibility, we provide the evaluation protocols and codes here.

The language dependency of /m/ in native Dutch and non-native English.

In forensic speaker comparisons, the current practice is to try to avoid comparisons between speech fragments in different languages. However, globalization requires an exploration of individual speech features that may show phonetic consistency across a speaker's languages. We predicted that the bilabial nasal /m/ may be minimally affected by the language spoken due to the involvement of the rigid nasal cavity in combination with a lack of fixed oral articulatory targets. The results show that indeed, L1 Dutch speakers (N = 53) had similar nasal formants and formant bandwidths when speaking in their L2 English as in their native language, suggesting language-independency of /m/ within speakers. In fact, acoustics seemed to rely more on the phonetic context than on the language spoken. Nevertheless, caution should still be exercised when sampling across languages when the languages' phoneme inventories and phonotactics show substantial differences.

Effects of sensorineural hearing loss on formant-frequency discrimination: Measurements and models

This study concerns the effect of hearing loss on discrimination of formant frequencies in vowels. In the response of the healthy ear to a harmonic sound, auditory-nerve (AN) rate functions fluctuate at the fundamental frequency, F0. Responses of inner-hair-cells (IHCs) tuned near spectral peaks are captured (or dominated) by a single harmonic, resulting in lower fluctuation depths than responses of IHCs tuned between spectral peaks. Therefore, the depth of neural fluctuations (NFs) varies along the tonotopic axis and encodes spectral peaks, including formant frequencies of vowels. This NF code is robust across a wide range of sound levels and in background noise. The NF profile is converted into a rate-place representation in the auditory midbrain, wherein neurons are sensitive to low-frequency fluctuations. The NF code is vulnerable to sensorineural hearing loss (SNHL) because capture depends upon saturation of IHCs, and thus the interaction of cochlear gain with IHC transduction. In this study, formant-frequency discrimination limens (DLFFs) were estimated for listeners with normal hearing or mild to moderate SNHL. The F0 was fixed at 100 Hz, and formant peaks were either aligned with harmonic frequencies or placed between harmonics. Formant peak frequencies were 600 and 2000 Hz, in the range of first and second formants of several vowels. The difficulty of the task was varied by changing formant bandwidth to modulate the contrast in the NF profile. Results were compared to predictions from model auditory-nerve and inferior colliculus (IC) neurons, with listeners’ audiograms used to individualize the AN model. Correlations between DLFFs, audiometric thresholds near the formant frequencies, age, and scores on the Quick speech-in-noise test are reported. SNHL had a strong effect on DLFF for the second formant frequency (F2), but relatively small effect on DLFF for the first formant (F1). The IC model appropriately predicted substantial threshold elevations for changes in F2 as a function of SNHL and little effect of SNHL on thresholds for changes in F1.

Comparing tools for automation of formant estimation in diverse pediatric populations

Pediatric vowel productions are notoriously hard to reliably quantify. Listener judgments and transcription are subjective, show low intra- and interrater reliability, and are influenced by listener bias. Formant estimation may offer more objectivity, but child speech is fraught with acoustic challenges. Children have less refined articulatory control and they make more production errors. They have high fundamental frequencies, wide formant bandwidths, more variable formant values, and increased subglottal coupling relative to adult speech. Historically, estimation of pediatric formants has been done manually, which is laborious and time-consuming. In recent years, automation tools have been developed to speed up the process. However, these tools have not been widely tested on pediatric speech samples. Even more critically, these tools have not been tested on diverse children. This study uses speech samples from children of color, children with disabilities, and children with both of these identities to compare three automation tools: SpeechMark® (Boyce etal., 2012), Fast Track™ (Barreda, 2021), and a custom Praat® (Boersma & Weenink, 2021) script written by the first author. Outcomes of each tool will be compared and contrasted. The discussion will review the considerations, benefits, and tradeoffs of each automation tool when working with diverse pediatric speech samples.

Formant-frequency discrimination in listeners with hearing loss: Behavioral results and neural models

In responses of the healthy ear to harmonic sounds, the amplitude of f0-related fluctuations of auditory-nerve (AN) responses varies along the tonotopic axis. Neural fluctuation amplitudes encode spectral peaks, including formant frequencies of vowels, because responses of inner-hair-cells (IHCs) tuned near spectral peaks are captured (or dominated) by a single harmonic, resulting in lower fluctuation amplitudes than IHCs tuned between spectral peaks. This code is robust across a wide range of sound levels and in background noise, and is converted into a rate-place representation in the auditory midbrain, where neurons are sensitive to low-frequency fluctuations. This neural-fluctuation code is vulnerable to sensorineural hearing loss (SNHL), because it depends upon saturation of IHCs and the interaction of cochlear amplification with IHC transduction. We tested formant-frequency discrimination in listeners with normal hearing or mild to moderate SNHL. The difficulty of the task was modulated by varying formant bandwidth. The relationship between spectral peaks and harmonic frequencies was also varied. Formant frequencies were 600 and 2000 Hz, in the range of first and second formants of many vowels. Results were compared to predictions made using model AN and midbrain neurons, with audiograms for the individual listeners included in the peripheral model. [Work supported by NIDCD-R01-001641]

Phonetic and phonological properties of the advanced tongue root feature in African languages

The paper presents an overview of the ATR (advanced tongue root) feature, which is fairly widespread among the Niger-Congo and Nilo-Saharan languages of the Macro-Sudan Belt. The phonetic section provides a description of the articulatory and acoustic correlates of the feature. The review of the articulatory basis of the feature is presented in a historical perspective, from the fi rst cineradiographic studies of the 1960s till the laryngoscopic studies of the recent years. The paper shows how the introduction of new methods shifted attention from the tongue root movement, which had been given primacy in production of the contrast, to aryepiglotto-epiglottal constriction and a synergistic system of laryngeal articulations. The paper provides a description of several acoustic measures that have been found to correlate with the ATR feature: the frequency of the fi rst formant, the fi rst formant bandwidth, the spectral center of gravity, the relative intensity of the fi rst to the second formant. Similarities and differences between ATR, height and tenseness, which have been largely discussed in the literature, are also described in the paper. The phonological section provides an overview of ATR harmony systems. The main attention is paid to the strong crosslinguistic tendency for [+ATR] to function as the dominant value in languages with ATR contrast among high vowels, but for [−ATR] to be dominant in languages in which ATR is only contrastive for non-high vowels. The paper describes diff erent manifestations of ATR dominance (assimilatory, allophonic, and coalescent dominance) among two major types of underlying vowel inventories.

Synthesis of Sung Spanish Vowels in Lyrical Singing by Sopranos

The aim of this paper is perform the synthesis of sung Spanish vowels considering the soprano vocal category of lyrical singers, including variation of sustained pitches with vibrato and tremolo effects, considering sounds from Spanish language. The Fant source-filter theory is used to model the production of the sung vowels: the source is based on the Rosenberg glottal pulse model and the filter (the vocal tract) is composed by an all-pole filter model with formant frequencies and bandwidths from the vowels of the Spanish language, obtained through experimental voice signals from two soprano singers. All the sounds synthesized are available to be accessed and they were submitted to a group of listeners which gave a very good evaluation with respect to the intelligibility and naturalness of the sounds.

Modal coupling of acoustic wave propagation and mechanical wall vibration in the time-varying vocal tract

Coupling between acoustic wave propagation and mechanical vibration of wall tissue along the vocal tract has long been considered to be an important loss mechanism in speech, affecting the frequencies and bandwidths of the lower formants. However, the detailed mechanism of energy transfer between fluid and structure as the shape of the vocal tract changes over time has not been fully elucidated. To examine this question, a time-domain finite-volume simulation of quasi-one-dimensional fluid flow and mechanical wall vibration was derived from underlying conservation laws. By calculating the eigenvalues and left and right eigenfunctions of the resulting implicit matrix recursion at each point in time, the sound field can be decomposed into modal vibrations describing coupled motion of air and tissue. Standing waves in the fluid within the vocal tract create sound waves that propagate in the wall tissue. Conversely, standing waves in the wall tissue excite waves in the fluid that propagate as sound. The coupling between airborne and mechanical modes of vibration and the transfer of energy between the two systems can easily be understood by examining the time-varying amplitude and phase of the spatial eigenfunctions. Illustrations are provided for simulations of VCV transitions.

Automated vowel space measurement of young children

Vowel formant frequencies are frequently reported as a measure of the articulatory workspace and have been used to study speech development and speech disorders across the lifespan. It is widely acknowledged that the measurement of F1/F2 frequencies for Vowel Space Area (VSA) estimation is a cumbersome task, wrought with challenges and uncertainties. Although there have been advancements in automated approaches, many of these tools continue to be challenged by child speech due to the anatomical differences and dynamic changes of the supraglottal, glottal, and subglottal regions of the airway during the years of speech development. This study examines automated VSA estimation in 4-year-old children using models that have been adapted to manage the higher fundamental frequencies, wider formant bandwidths, and higher amplitudes of subglottal resonances detected in child speech. The speech of six, native American English-speaking, four-year-old children with no history of hearing or speech disorders were included in this investigation. 48 words, including eight American English vowels, were analyzed using traditional manual methods and the SpeechMark MATLAB toolbox automated VSA estimator. The toolbox automates the generation of a convex hull enclosing the child VSA and computation of the area while accounting for the anatomical/acoustic differences that occur during early speech development.

Are Acoustic Markers of Voice and Speech Signals Affected by Nose-and-Mouth-Covering Respiratory Protective Masks?

BackgroundWorldwide use of nose-and-mouth-covering respiratory protective mask (RPM) has become ubiquitous during COVID19 pandemic. Consequences of wearing RPMs, especially regarding perception and production of spoken communication, are gradually emerging. The present study explored how three prevalent RPMs affect various speech and voice sound properties.MethodsPre-recorded sustained [a] vowels and read sentences from 47 subjects were played by a speech production model (‘Voice Emitted by Spare Parts’, or ‘VESPA’) in four conditions: without RPM (C1), with disposable surgical mask (C2), with FFP2 mask (C3), and with transparent plastic mask (C4). Differences between C1 and masked conditions were assessed with Dunnett's t test in 26 speech sound properties related to voice production (fundamental frequency, sound intensity level), voice quality (jitter percent, shimmer percent, harmonics-to-noise ratio, smoothed cepstral peak prominence, Acoustic Voice Quality Index), articulation and resonance (first and second formant frequencies, first and second formant bandwidths, spectral center of gravity, spectral standard deviation, spectral skewness, spectral kurtosis, spectral slope, and spectral energy in ten 1-kHz bands from 0 to 10 kHz).ResultsC2, C3, and C4 significantly affected 10, 15, and 19 of the acoustic speech markers, respectively. Furthermore, absolute differences between unmasked and masked conditions were largest for C4 and smallest for C2.ConclusionsAll RPMs influenced more or less speech sound properties. However, this influence was least for surgical RPMs and most for plastic RPMs. Surgical RPMs are therefore preferred when spoken communication is priority next to respiratory protection.

Fast Track: fast (nearly) automatic formant-tracking using Praat

Abstract Fast Track is a formant tracker implemented in Praat that attempts to automatically select the best analysis from a set of candidates. The best track is selected by modeling smooth formant contours across the entirety of the sound, providing the researcher with rich information about static and dynamic formant properties. Fast Track returns text files containing acoustic information (formant frequencies, formant bandwidths, fundamental frequency, etc.) sampled every 2 ms, generates images showing the winning analysis and comparing alternate analyses, and creates log files detailing analysis information for each file. Fast Track features a modular workflow that allows for analysis steps to be run (and re-run) independently as necessary, and is designed to allow for easy correction of tracking errors by allowing the user to override the automatic analysis, or manually edit tracks where necessary. In addition, Fast Track includes tools to aggregate data across tokens, and to easily create vowel plots of mean values or time-varying formant contours. The design and use of Fast Track are outlined using a re-analysis of the Hillenbrand et al. (1995) dataset, which suggests that Fast Track can be very accurate in cases where signal properties allow for reliable formant estimates.

The Formant Bandwidth as a Measure of Vowel Intelligibility in Dysphonic Speech

The current paper examined the impact of dysphonia on the bandwidth of the first two formants of vowels, and the relationship between the formant bandwidth and vowel intelligibility. Speaker participants of the study were 10 adult females with healthy voice and 10 adult females with dysphonic voice. Eleven vowels in American English were recorded in /h/-vowel-/d/ format. The vowels were presented to 10 native speakers of American English with normal hearing, who were asked to select a vowel they heard from a list of /h/-vowel-/d/ words. The vowels were acoustically analyzed to measure the bandwidth of the first and second formants (B1 and B2). Separate Wilcoxon rank sum tests were conducted for each vowel for normal and dysphonic speech because the differences in B1 and B2 were found to not be normally distributed. Spearman correlation tests were conducted to evaluate the association between the difference in formant bandwidths and vowel intelligibility between the healthy and dysphonic speakers. B1 was significantly greater in dysphonic vowels for seven of the eleven vowels, and lesser for only one of the vowels. There was no statistically significant difference in B2 between the normal and dysphonic vowels, except for the vowel /i/. The difference in B1 between normal and dysphonic vowels strongly predicted the intelligibility difference. Dysphonia significantly affects B1, and the difference in B1 may serve as an acoustic marker for the intelligibility reduction in dysphonic vowels. This acoustic-perceptual relationship should be confirmed by a larger-scale study in the future.

The Impact Of Lexical Stress Type and Position of a Vowel Within a Phrase on Vowel Formant Bandwidths

The article sets out to explore the impact of different types of lexical stress and word position within a phrase as well as the interaction of these factors on formant bandwidths. The findings contribute to establishing the role of these formant features as acoustic correlates of the Russian lexical stress. The experimental material for analysis is presented by the Russian sound [a] embedded in a word in a natural language carrier phrase. The word position is changed from phrase initial to the phrase final one and in each position the target word is uttered with a neutral and an emphatic stress by four speakers. The Praat software is used to extract the mean values of the first four formant bandwidths of the target vowel. Two-way ANOVA is carried out to establish the significance of difference between neutrally and emphatically stressed vowels in all the three phrasal positions. The impact of phrasal position and stress type is clearly pronounced and is valid for all the four speakers. Specific trends in bandwidth alterations are hard to identify in most cases because of inconsistent fluctuations of formant bandwidths and a heavy influence of the speakers' idiolects: formant bandwidths expand, contract or remain unchanged. An explanation of the results obtained is suggested.

Bioacoustic detection of Callosobruchus chinensis and Callosobruchus maculatus in bulk stored chickpea (Cicer arietinum) and green gram (Vigna radiata)

Insects are one of the major reasons for post-harvest losses in stored food legumes. Real time detection at an early stage of insect infestation can overcome the foremost problem of infestation. Acoustic detection method detects the hidden as well as moving insects by amplifying and filtering their motility and feeding sounds. Crawling and feeding activities of two insects Callosobruchus chinensis and Callosobruchus maculatus in chickpea (Cicer arietinum) and green gram (Vigna radiata) were detected by using an insect detection probe placed in an acoustically insulated bin. Temperature and relative humidity were also measured through sensors mounted on the same probe. Captured sounds were pre-screened and analysed by using sound analysis software. Significant differences were found between the amplitude of two insects in both the legumes. Amplitudes of Callosobruchus chinensis were 79.32 dB and 84.01 dB for 59 ms duration in chickpea and green gram, respectively, whereas, the amplitude of Callosobruchus maculatus was 97.65 dB and 95.53 dB for 68 ms in chickpea and green gram, respectively. The detection range of the acoustic sensor was 300 mm omnidirectional. Formants (F1, F2, F3 and F4), formant bandwidth (FBW3), frequency and spectral power were observed as principal components for the detection of insects’ infestation (C. chinensis and C. maculatus) in bulk stored food legumes (chickpea and green gram). Formants of sound spectra found a specific role in the discrimination of insect and background noise. The findings of this study elucidated that acoustic detection method will provide rapid and non-destructive detection of infestation in bulk stored food legumes.

Analisis Pola Karakteristik Suara Pada Kasus Forensik

ABSTRACTAnalysis of voice characteristic pattern has an important role in the forensic science. It can apply to solve criminal cases in order to straighten justice and law in the Criminal Justice System. This study tries to analyze the pattern of voice print by comparing the evidence of recorded sound from the mobile phone with the sound of testee to determine their equality or difference. In answering the problem stated in the research hypothesis, then used comprehensive analysis based on voice acoustical characteristics, e. g. : spectrogram analysis, pitch statistical analysis, formant and formant bandwidth statistical analysis. Based on statistical test criteria with level of significance α = 0,05 reveals that the sound of evidence and the sound of testee are equal. Those sounds come from one person, in such a way that we can determine a suspect. Key words : Voice Analysis, Forensic Science, Criminal Justice System.

Measuring communication difficulty through effortful speech production during conversation

This study describes the use of a novel conversation elicitation framework to collect fluent, dynamic conversational speech in simulated realistic acoustic environments of varying complexities. Our aim is to quantify speech modifications during conversation, which characterize effortful speech, as a function of the difficulty of the acoustic environment. We report speech production data at the acoustic-phonetic level (vocal level, mid-frequency emphasis, formant frequencies and formant bandwidths), as well as at higher levels of analysis including utterance duration and turn overlap durations. The sensitivity and test-retest reliability of different speech production measures to changes in acoustic environment are reported. We propose a multi-dimensional view of effortful speech modifications. Considering speech modifications across different linguistic levels provides a richer view of the effects of the acoustic environment on communication as compared with consideration of low-level acoustic-phonetic markers alone. Finally, we describe how consideration of speech modification data may form the basis of a measure of communication effort with scope for the assessment of the impacts of hearing impairment and amplification upon ease of spoken communication.

Acoustic cues to linguistic profiling? Machine learning of phonetic features of African American English

Listeners can readily differentiate words spoken in an African American English (AAE) dialect from a Standard American English (SAE) dialect, even in absence of distinctive morphosyntactic features. However, it is still unclear what acoustic-phonetic cues listeners utilize to rapidly distinguish AAE from SAE. This study investigates the informativeness of various acoustic-phonetic cues to the characterization of AAE dialect. V and VC sequences (with C = /n/, /m/, /l/, /r/) from speech of 7 female speakers (4 SAE and 3 AAE), recorded during sociolinguistic interviews, were randomly selected and acoustically analyzed, controlling for coarticulatory context. Acoustic cues of F1, F2, F3, F4 formant trajectories, formant bandwidth, pitch variation, duration, intensity, and voice quality measures (e.g., harmonic-to-noise ratio, jitter, shimmer, and spectral slope) were measured in these segments to identify the extent of their contribution to separating AAE and SAE. The results from machine learning modeling of acoustic cues demonstrate that speech within an AAE dialect entails distinct acoustic characteristics and voice quality compared to SAE speech. These separate acoustic patterns between AAE and SAE dialect indicate the need for including dialect-specific acoustic cues both in automatic speech recognition applications and clinical assessments of speech-language disorders.

Performance comparison of multitaper techniques for speaker verification with expressive speech

In this paper, we provide a comparative study of spectral front-end features used as representations for speech signals by processing multitaper magnitude and phase spectra, for speaker verification with expressive speech. In particular, the multitaper modified group delay function (MT-MOGDF) and multitaper magnitude (MT-MAG) spectra of the speech signals are employed to obtain low variance estimates of speech spectra. We observe that the cues that aid in representation of expressive speech are evident in the MT-MOGDF spectrum than the MT-MAG spectrum in terms of mean Formant value and Formant bandwidth. Our extensive experimental study on a speaker verification system with a Gaussian mixture model based universal background model classifier on expressive speech using the IITKGP-SESC and EMODB databases show that MT-MOGDF performs better than MT-MAG technique, in terms of equal error rate and minimum decision cost function. This improvement due to MT-MOGDF is owed to a better representation and a low-variance estimate of the speech spectrum. Our results highlight the utility of MT-MOGDF as a potential alternative for MT-MAG representation for speaker verification problems in general.

Predicting intelligibility of dysphonic speech with automatic measurement of vowel related parameters

Speakers with dysphonia frequently report difficulty with being understood in noisy places. Perception of vowels plays an important role in intelligibility. Dysphonic speech is typically characterized both by shorter periodic intervals and an increased proportion of noise vs. harmonic components, potentially obscuring acoustic cues for identification of vowels. This study examined whether vowel-related acoustic measures would predict intelligibility of dysphonic speech. Sentences from the Consensus Auditory-Perceptual Evaluation of Voice (CAPE-V) were recorded from 18 dysphonic speakers (6 adult females, 6 adult males, and 6 children) and 3 normal speakers (1 adult female, 1 adult male, and 1 child). These sentences were analyzed acoustically by SpeechMark® Wavesurfer plug-in version, a speech analysis tool based on the landmark theory of speech perception. The sentences were presented to 45 listeners for perceptual rating of intelligibility at 3 S/N ratios. The following output parameters were compared: vowel space area, number of vowels whose F1 and F2 values fell within a standard vowel quadrilateral (defined based on existing literature), and number of vowels with normal formant bandwidth. Preliminary results indicated that number of vowels with normal formant bandwidth falling within the quadrilateral significantly predicted the intelligibility ratings of dysphonic speech at all noise levels.

New Evidence That Nonlinear Source-Filter Coupling Affects Harmonic Intensity and fo Stability During Instances of Harmonics Crossing Formants

The traditional source-filter theory of voice production describes a linear relationship between the source (glottal flow pulse) and the filter (vocal tract). Such a linear relationship does not allow for nor explain how changes in the filter may impact the stability and regularity of the source. The objective of this experiment was to examine what effect unpredictable changes to vocal tract dimensions could have on fo stability and individual harmonic intensities in situations in which low frequency harmonics cross formants in a fundamental frequency glide. To determine these effects, eight human subjects (five male, three female) were recorded producing fo glides while their vocal tracts were artificially lengthened by a section of vinyl tubing inserted into the mouth. It was hypothesized that if the source and filter operated as a purely linear system, harmonic intensities would increase and decrease at nearly the same rates as they passed through a formant bandwidth, resulting in a relatively symmetric peak on an intensity-time contour. Additionally, fo stability should not be predictably perturbed by formant/harmonic crossings in a linear system. Acoustic analysis of these recordings, however, revealed that harmonic intensity peaks were asymmetric in 76% of cases, and that 85% of fo instabilities aligned with a crossing of one of the first four harmonics with the first three formants. These results provide further evidence that nonlinear dynamics in the source-filter relationship can impact fo stability as well as harmonic intensities as harmonics cross through formant bandwidths.