Evaluating acoustic speaker normalization algorithms: evidence from longitudinal child data

Evaluating normalization accounts against the dense vowel space of Central Swedish

Talkers vary in the phonetic realization of their vowels. One influential hypothesis holds that listeners overcome this inter-talker variability through pre-linguistic auditory mechanisms that normalize the acoustic or phonetic cues that form the input to speech recognition. Dozens of competing normalization accounts exist-including both accounts specific to vowel perception and general purpose accounts that can be applied to any type of cue. We add to the cross-linguistic literature on this matter by comparing normalization accounts against a new phonetically annotated vowel database of Swedish, a language with a particularly dense vowel inventory of 21 vowels differing in quality and quantity. We evaluate normalization accounts on how they differ in predicted consequences for perception. The results indicate that the best performing accounts either center or standardize formants by talker. The study also suggests that general purpose accounts perform as well as vowel-specific accounts, and that vowel normalization operates in both temporal and spectral domains.

Spanish in Albuquerque, New Mexico: Spanish-English Bilingual Adults’ and Children’s Vocalic Realizations

This study explores vocalic production and variation in 29 Spanish-English bilingual children and adults from Albuquerque, New Mexico. Linear mixed-effects models analyzed the effects of lexical stress, word position, phonetic context, Spanish use, and lexical frequency on the F1 (height) and F2 (frontedness) values of 2041 /i e a o u/ vowels. Importantly, results show that /u/ fronting is pervasive in both children and adults’ speech, but in contrast to adults’ more general /u/ fronting, children’s Spanish use also predicted atonic /u/ fronting. Expanding the range of data to include children also showed that children’s realizations displayed a generalized stress effect, whereby the atonic space was condensed compared to the tonic space. The generalized stress effect was absent among the adults. Changes in the degree of phonetic convergence between the adults and children are attributed to acquisitional paths and demographic changes in their community.

The Entanglement of Dialectal Variation and Speaker Normalization

This paper explores the relationship between speaker normalization and dialectal identity in sociolinguistic data, examining a database of vowel formants collected from 88 monolingual American English speakers in Michigan's Upper Peninsula. Audio recordings of Finnish- and Italian-heritage American English speakers reading a passage and a word list were normalized using two normalization procedures. These algorithms are based on different concepts of normalization: Lobanov, which models normalization as based on experience with individual talkers, and Labov ANAE, which models normalization as based on experience with scale-factors inherent in acoustic resonators of all kinds. The two procedures yielded different results; while the Labov ANAE method reveals a cluster shifting of low and back vowels that correlated with heritage, the Lobanov procedure seems to eliminate this sociolinguistic variation. The difference between the two procedures lies in how they treat relations between formant changes, suggesting that dimensions of variation in the vowel space may be treated differently by different normalization procedures, raising the question of how anatomical variation and dialectal variation interact in the real world. The structure of the sociolinguistic effects found with the Labov ANAE normalized data, but not in the Lobanov normalized data, suggest that the Lobanov normalization does over-normalize formant measures and remove sociolinguistically relevant information.

A regression approach to vowel normalization for missing and unbalanced data

Researchers investigating the vowel systems of languages or dialects frequently employ normalization methods to minimize between-speaker variability in formant patterns while preserving between-phoneme separation and (socio-)dialectal variation. Here two methods are considered: log-mean and Lobanov normalization. Although both of these methods express formants in a speaker-dependent space, the methods differ in their complexity and in their implied models of human vowel-perception. Typical implementations of these methods rely on balanced data across speakers so that researchers may have to reduce the data available in the analyses in missing-data situations. Here, an alternative method is proposed for the normalization of vowels using the log-mean method in a linear-regression framework. The performance of the traditional approaches to log-mean and Lobanov normalization against the regression approach to the log-mean method using naturalistic, simulated vowel-data was investigated. The results indicate that the Lobanov method likely removes legitimate linguistic variation from vowel data and often provides very noisy estimates of the actual vowel quality associated with individual tokens. The authors further argue that the Lobanov method is too complex to represent a plausible model of human vowel perception, and so is unlikely to provide results that reflect the true perceptual organization of linguistic data.

Methods for eliciting, annotating, and analyzing databases for child speech development

Methods from automatic speech recognition (ASR), such as segmentation and forced alignment, have facilitated the rapid annotation and analysis of very large adult speech databases and databases of caregiver-infant interaction, enabling advances in speech science that were unimaginable just a few decades ago. This paper centers on two main problems that must be addressed in order to have analogous resources for developing and exploiting databases of young children's speech. The first problem is to understand and appreciate the differences between adult and child speech that cause ASR models developed for adult speech to fail when applied to child speech. These differences include the fact that children's vocal tracts are smaller than those of adult males and also changing rapidly in size and shape over the course of development, leading to between-talker variability across age groups that dwarfs the between-talker differences between adult men and women. Moreover, children do not achieve fully adult-like speech motor control until they are young adults, and their vocabularies and phonological proficiency are developing as well, leading to considerably more within-talker variability as well as more between-talker variability. The second problem then is to determine what annotation schemas and analysis techniques can most usefully capture relevant aspects of this variability. Indeed, standard acoustic characterizations applied to child speech reveal that adult-centered annotation schemas fail to capture phenomena such as the emergence of covert contrasts in children's developing phonological systems, while also revealing children's nonuniform progression toward community speech norms as they acquire the phonological systems of their native languages. Both problems point to the need for more basic research into the growth and development of the articulatory system (as well as of the lexicon and phonological system) that is oriented explicitly toward the construction of age-appropriate computational models.

Fronting of Southern British English high-back vowels in articulation and acoustics

The fronting of the two high-back vowels /uː/ and /ʊ/ in Southern British English is very well documented, but mainly in the acoustic domain. This paper presents articulatory (ultrasound) data, comparing the relative tongue position of these vowels in fronting and non-fronting consonantal contexts, i.e., preceding a coronal consonant (food, foot) and preceding a coda /l/ (fool, full). Particular attention is paid to the comparison between articulatory results and corresponding acoustic measurements of F2 in both vowels. Results show that the average differences between food and foot and their dynamic profiles are similar in articulation and acoustics. In /uːl/ sequences (fool), tongue position is more advanced than could be inferred from its low F2. In addition, even though the tongue position in fool and full is clearly distinct, there is no comparable corresponding difference in F2. This suggests that the common articulatory metaphor that characterises F2 increase as fronting must be used cautiously. In the case of English high-back vowel fronting, special attention must be paid to the flanking consonants when estimating vowel distances. This paper also provides specific recommendations for recording and analysing ultrasound data in research on vowel variation and change.

An ultrasound study of Canadian French rhotic vowels with polar smoothing spline comparisons

This is an acoustic and articulatory study of Canadian French rhotic vowels, i.e., mid front rounded vowels /ø œ̃ œ/ produced with a rhotic perceptual quality, much like English [ɚ] or [ɹ], leading heureux, commun, and docteur to sound like [ɚʁɚ], [kɔmɚ̃], and [dɔktaɹʁ]. Ultrasound, video, and acoustic data from 23 Canadian French speakers are analyzed using several measures of mid-sagittal tongue contours, showing that the low F3 of rhotic vowels is achieved using bunched and retroflex tongue postures and that the articulatory-acoustic mapping of F1 and F2 are rearranged in systems with rhotic vowels. A subset of speakers' French vowels are compared with their English [ɹ]/[ɚ], revealing that the French vowels are consistently less extreme in low F3 and its articulatory correlates, even for the most rhotic speakers. Polar coordinates are proposed as a replacement for Cartesian coordinates in calculating smoothing spline comparisons of mid-sagittal tongue shapes, because they enable comparisons to be roughly perpendicular to the tongue surface, which is critical for comparisons involving tongue root position but appropriate for all comparisons involving mid-sagittal tongue contours.