SingaKids-Mandarin: Speech Corpus of Singaporean Children Speaking Mandarin Chinese

Abstract

We present SingaKids-Mandarin, a speech corpus of 255 Singaporean children aged 7 to 12 reading Mandarin Chinese, for a total of 125 hours of data (75 hours of speech) and 79,843 utterances. This corpus is phonetically balanced and detailed in human annotations, including phonetic transcriptions, lexical tone markings, and proficiency scoring at the utterance level. The reading scripts span a diverse set of utterance styles, covering syllable-level minimal pairs, words, phrases, sentences, and short stories. We analyze the acoustic properties of Singaporean children. We also observe that while the lack of the neutral tone is the same for Singaporean adults and children, the phonetic pronunciation patterns in these two age groups differ: although Singaporean adults tend to front their retroflex, nasal, and palatal consonants, Singaporean children show both fronting and backing in these consonants. For future work, we plan to develop computer-assisted pronunciation training (CAPT) systems with SingaKids-Mandarin.

L2-ARCTIC: A Non-native English Speech Corpus

“…The ISLE Speech Corpus [21] contains mispronunciation tags and is open for academic access, but it only focuses on a limited group of English learners (German and Italian). SingaKids-Mandarin [22] has a rich set of speech data, but it only focuses on mispronunciation patterns in Singapore children's Mandarin speech. In fact, most existing mispronunciation detection systems use their private datasets, which makes it difficult to compare experimental results across different publications [19,[23][24][25].…”

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

“…This is an efficient way to record phonetically balanced materials from a large number of school-aged children. Examples of databases that use this method are the TIDIGITS corpus (Leonard, 1984), the CMU KIDS corpus (Eskenazi, 1996; Eskenazi et al, 1997), the CID corpus (Miller et al, 1996; Lee et al, 1999), and the OGI / CSLU Kid's Speech corpus (Shobaki et al, 2000, 2007) for American English; the ChildIt corpus for Italian (Giuliani et al, 2006; Gerosa et al, 2007); the Swedish portion of the PF STAR Children's Speech Corpus (Batliner et al, 2005); and the SingaKids–Mandarin corpus for Singapore Mandarin Chinese (Chen et al, 2016). …”

Multi-Task Learning for Mispronunciation Detection on Singapore Children’s Mandarin Speech

“…Speech productions of Singapore children [1] possess even more variability due to the multilingual environment in the citystate, causing inter-influences from Chinese languages (e.g., Hokkien and Cantonese), English dialects (e.g., American and British), Malay languages (e.g., Bahasa Melayu and Bahasa Indonesia) and Indian languages (e.g., Hindi and Tamil). For example, retroflex fricatives and affricates (e.g., /SH/ in Pinyin) in Mandarin might be influenced from Hokkien, where the retroflex place of articulation is not phonemic.…”

“…Data collection on children's speech is much more difficult than that of adults, as children usually have limited attention span and vocabulary, especially those at young ages. Another reason is the high acoustic and linguistic variabilities in children's speech: acoustic characteristics varies across age and gender [1,2], there are larger variances in speaking rate and vocal effort [3] in children's speech, and children have immature command of word usage, grammar, and linguistic structure [4].…”

“…Specificity in this work, the acoustic characteristics of adults and children are jointly learned through shared hidden layers of a deep neural network, yet they still retain their respective targets using two distinct softmax layers. (1). There are five tones in Mandarin: Tone 1 to Tone 5, in which Tone 5 is neutral and has no specific pitch contour, it is analogous to an unstressed syllable in English.…”