Learning Mechanisms in Early Vocal Development

标准的发音阶段(平均7个月)的完成是一个重要的里程碑：在这个年龄段，婴儿开始产生辅音-元音音节，这是口语的基本组成部分。来自我的实验室的研究涉及不同程度听力损失的婴儿，研究表明，成年说话者的听力对于在生命的第一年实现类似言语的喋喋不休是必不可少的。此外，学习英语、法语和阿拉伯语的婴儿，学会了以微妙的不同方式喋喋不休。宝宝是如何学会用母语发音的？我建议测试一个言语运动学习模型，在这个模型中，婴儿学习从听到的语音中感知言语类别，并试图在喋喋不休时产生这些类别。婴儿的动机是减少内化的知觉类别与在喋喋不休过程中产生的言语之间的匹配错误。成人照顾者的言语输入(A)提供有关母语言语类别的信息，(B)鼓励发声探索，(C)加强言语练习。这一模型可以与另一种模型形成对比，在另一种模型中，当成年人模仿类似于成人语言中的言语类别的婴儿言语时，类似言语的婴儿话语被强化。这个研究项目的一个目的是描述成人和他们的婴儿在三种背景下的模仿模式：(1)听力正常的婴儿和他们说英语的母亲；(2)听力正常的婴儿和他们的阿拉伯语母亲；(3)听力受损的婴儿，植入人工耳蜗，与他们的阿拉伯语母亲互动。我们预计，母体模仿的发生频率将比目前一些语言发育计算机模型预测的要少得多，并发现母亲会根据婴儿输出的意义调整她们的言语反应。在第二项研究中，我们将教婴儿通过接触外语来产生外语元音，但不模仿或扩大婴儿的发音。将描述婴儿对外语语音类别的感知和产生的变化，并与对照组进行比较。我们假设研究结果将支持内在动机和自我监督学习作为支持语音学习的机制。了解语音产生是如何发展的，在语音治疗和减少口音等领域一直都有重要的应用。最近，婴儿语音产生研究已被应用于语音识别和合成、机器学习和机器人学等领域。目前正在通过机器学习模型对理论账户进行测试，而没有足够的人力数据来指导其实施或验证其产出。这项研究计划将产生大量详细的语音和声学数据样本，描述不同条件下婴儿的声音输出。

Achievement of the canonical babbling stage (on average at 7 months) is an important milestone: at this age the infant begins to produce consonant-vowel syllables that are the fundamental building blocks of spoken language. Research from my lab, involving infants with varying levels of hearing loss, shows that hearing speech from adult talkers is essential to achieving speech-like babble in the first year of life. Furthermore, babies learning English, French, and Arabic, learn to babble in subtly different ways. How do babies learn to babble with native language sound characteristics? I propose to test a model of speech motor learning in which the infant learns to perceive speech categories from heard speech and attempts to produce those categories when babbling. The baby is motivated to reduce errors in the match between internalized perceptual categories and speech produced during babble. Speech input from adult caregivers (a) provides information about native-language speech categories, (b) encourages vocal exploration, and (c) reinforces speech practice. This model can be contrasted with an alternative model in which speech-like infant utterances are reinforced when adults mimic infant speech that is similar to speech categories in the adult language. One aim of this research program is to describe patterns of mimicry between adults and their infants in three contexts: (1) normal hearing infants and their English-speaking mothers; (2) normal hearing infants and their Arabic speaking mothers; (3) hearing impaired infants, with cochlear implants, in interaction with their Arabic speaking mothers. We expect that maternal mimicry will occur much less frequently than is predicted by some current computer models of speech development and to find that mothers adjust their speech responses to the meaningfulness of their infant's outputs. In a second study we will teach infants to produce foreign language vowels by providing exposure to the foreign language, but without imitating or expanding infant utterances. Changes in the infants' perception and production of foreign language sound categories will be described in comparison to a control group. We hypothesize that the results will support intrinsic motivation and self-supervised learning as mechanisms that support speech learning. Understanding how speech production develops has always had important applications in fields such as speech therapy and accent reduction. More recently, infant speech production research has been applied in the areas of speech recognition and synthesis, machine learning and robotics. Theoretical accounts are currently being tested via machine learning models without the benefit of sufficient human data to guide their implementation or validate their outputs. This research program will produce large samples of detailed phonetic and acoustic data describing infant vocal output under varying conditions.