Sofia Strömbergsson

Children’s perception of their synthetically corrected speech production

Abstract We explore children’s perception of their own speech – in its online form, in its recorded form, and in synthetically modified forms. Children with phonological disorder (PD) and children with typical speech and language development (TD) performed tasks of evaluating accuracy of the different types of speech stimuli, either immediately after having produced the utterance or after a delay. In addition, they performed a task designed to assess their ability to detect synthetic modification. Both groups showed high performance in tasks involving evaluation of other children’s speech, whereas in tasks of evaluating one’s own speech, the children with PD were less accurate than their TD peers. The children with PD were less sensitive to misproductions in immediate conjunction with their production of an utterance, and more accurate after a delay. Within-category modification often passed undetected, indicating a satisfactory quality of the generated speech. Potential clinical benefits of using corrective re-synthesis are discussed.

Children's recognition of their own recorded voice: influence of age and phonological impairment.

Children with phonological impairment (PI) often have difficulties perceiving insufficiencies in their own speech. The use of recordings has been suggested as a way of directing the childs attention toward his/her own speech, despite a lack of evidence that children actually recognize their recorded voice as their own. We present two studies of childrens self-voice identification, one exploring developmental aspects, and one exploring potential effects of having a PI. The results indicate that children from 4 to 8 years recognize their recorded voice well (around 80% accuracy), regardless of whether they have a PI or not. A subtle change in this ability from 4 to 8 years is observed that could be linked to a development in short-term memory. Clinically, one can indeed expect an advantage of using recordings in therapy; this could constitute an intermediate step toward the more challenging task of online self-monitoring.

Today's Most Frequently Used F 0 Estimation Methods, and Their Accuracy in Estimating Male and Female Pitch in Clean Speech.

Variation in fundamental frequency (F0) constitutes a valuable source of information for researches across many disciplines, with a shared interest in speech. Different methods for estimating F0 vary in estimation accuracy and accessibility, and there is yet no gold standard. Through a bibliometric survey, this study examines what methods were the most frequently used in the speech scientific community during the years 20102016. Secondly, the most used methods are evaluated against a ground truth reference, with a specific focus on their accuracy in estimating F0 in male and female speakers, respectively. The results show that Praat is the dominant method by far, followed by STRAIGHT, RAPT and YIN. This pattern holds across a range of different research areas, although within Acoustics and Engineering, Praat’s dominance is less pronounced. In the evaluation including Praat, RAPT and YIN – with their default and gender-adapted settings – Praat also proved to be the most accurate. The finding that adapting Praat’s pitch range settings by gender leads to further improvements should encourage researchers to do this routinely.

Approximating phonotactic input in children’s linguistic environments from orthographic transcripts

This paper presents a new speaker change detection system based on Long Short-Term Memory (LSTM) neural networks using acoustic data and linguistic content. Language modelling is combined with two different Joint Factor Analysis (JFA) acoustic approaches: i-vectors and speaker factors. Both of them are compared with a baseline algorithm that uses cosine distance to detect speaker turn changes. LSTM neural networks with both linguistic and acoustic features have been able to produce a robust speaker segmentation. The experimental results show that our proposal clearly outperforms the baseline system.Child-directed spoken data is the ideal source of support for claims about children’s linguistic environments. However, phonological transcriptions of child-directed speech are scarce,compared to s ...